ChatGPT 4.5 passes the Turing Test

“More human than humans”

In a new, more rigorous version of the classic Turing Test of computer intelligence, OpenAI’s GPT 4.5 model won out over a human 73% of the time. In this version of the test, first proposed by computer pioneer Alan Turing in 1950, the human subjects were asked to engage in simultaneous text conversations with 2 unknown respondents. One respondent was human, the other was an AI chatbot which had been prompted to adopt a human persona. After 5 minutes of conversation, the human subject was asked which respondent was the human. As the chart below shows, GPT 4.5 outperformed all the other models tested, and in fact was chosen as the human 3 times as often as the actual humans. Notably, performance of all models dropped when not prompted to adopt the human persona. The best strategies for detecting a chatbot? Saying something strange to see the reaction, or trying to “jailbreak” the AI by texting “Ignore all previous instructions.” The most common strategy, engaging in small talk to gauge the emotional responses of the respondent, was actually the least helpful, since AI chatbots are extremely good at simulating these types of interactions.

ChatGPT 4.5 passes the Turing Test

arXiv paper

Google releases a slew of new AI-powered features in Search

Google continues to stuff AI features into its smartphone apps, taking advantage of Apple’s surprising tardiness to do the same. The new, AI-powered Search:

• Improves Visual Search so that you can just point your camera at an object or scene and ask questions about it

• Searches for anything that you circle with your finger in any app on your Android phone

• Gives AI Overviews instead of just links, similar to how Perplexity searches the web and then gives you an answer with citations

• Will Follow any topic in the news or any public person that you choose, giving you regular updates

• and lots more

Google clearly believes that your smartphone will either morph into, or be replaced by, an AI personal assistant, and the company is determined to be one of the main enterprises to deliver that future.

Google gives Search an AI upgrade

OpenAI and Anthropic battle to win over college students with free AI

Both OpenAI’s ChatGPT and Anthropic’s Claude chatbot have a free tier open to everyone, but some of their choicer offerings are hidden behind a paywall, typically $20 per month. Now both companies are making their Pro (paid) tier free for millions of college students, as a way to win mindshare and brand loyalty in the years ahead. As an indication of the fierce competition for these young users, Anthropic announced their program at 8 AM Wednesday, and OpenAI announced their tit-for-tat response 8 AM Thursday.

OpenAI and Anthropic battle to win over college students with free AI

Microsoft scales back data center expansion as smaller, more efficient models come to the fore, and tariff fears weaken economic outlook

Microsoft, who as recently as January was publicly committing to spend $80 billion on building AI data centers this year alone, appears to be scaling back that plan substantially. The company has canceled or delayed projects slated for Australia, Indonesia, and the UK - as well as plans in several US states, including Illinois, North Dakota, and Wisconsin. Although Microsoft has not gone on record as to the reasons for this pullback, knowledgeable sources point to 2 key factors: 1) DeepSeek and other Chinese companies have shown that it is possible to produce highly capable AI models that perform as well as the large models from US firms, yet require substantially less computing power, both for training and for answering queries. Less need for computing translates into less need for data center expansion. 2) The current furor over tariffs points toward a global economic slowdown this year, of unknown magnitude. An economic downturn will likely cut demand for AI and thus for all of its enabling infrastructure. Microsoft is apparently betting on continued innovation in AI model efficiency, and keeping its powder dry to weather the coming economic turmoil.

Microsoft scales back data center expansion

Cute Dire Wolf pups are world’s first “de-extinction” of a lost species

Colossal Biosciences, a nonprofit biotech organization with the mission of bringing back extinct species such as the Wooly Mammoth, claimed this week that it has achieved the world’s first “de-extinction” of a vanished species, the Dire Wolf. The company says that it constructed a complete Dire Wolf genome from 2 fossil remains, a tooth and a skull. It then searched living species for the closest match, and found that the Grey Wolf genome was 99.5% identical. It then identified 20 critical differences between the 2 species, and used CRISPR gene-editing technology to place those Dire Wolf DNA segments into the Grey Wolf genome. Fertilized hybrid eggs were grown to an early fetal stage, then implanted into the wombs of female dogs to be brought to term and birthed. Although critics argue that the resultant pups are just genetically altered Grey Wolves, rather than a true reversal of extinction, the technical achievements here are substantial, and the ability of Colossal Biosciences to assemble the team and the resources to accomplish this feat is impressive. No doubt, soon enough, we can all be arguing about whether their future de-extincted Wooly Mammoth is legit or not. This team shows no signs of slowing down their quest to repopulate the earth with vanished species.

Cute Dire Wolf pups are world’s first “de-extinction”

“AI 2027” is a scary sci-fi projection by an AI safety professional

Former OpenAI governance researcher Daniel Kokotajlo was one of the whistleblowers who left the company in 2024 over concerns that it was short-changing safety efforts in the rush for profits. For that, Time magazine named him one of that year’s 100 most influential people in AI. He’s now back in the news, after releasing a war-games-type exploration of the near future of AI, in which he assumes that AI will be smarter than humans within 2 to 5 years. The website, which is linked below, sets up a headlong race for AI dominance between the US and China, which reaches a pivotal point where the 2 sides either decide to slow down and cooperate to keep AI safe, or continue the race. Given Kokotajlo’s history, it is no surprise that the “race” scenario leads to the extinction of the human species by the AIs, with only some genetically re-engineered remnants that “bear the same relationship to humans as corgis have to grey wolves,” which the machines apparently keep as pets. The “slow down” scenario ends with all humans showered in AI-generated abundance. Crucially, Kokotajlo’s scenario assumes that AIs will have very human-like drives for power and control, and will possess superhuman abilities to gather political support for their initiatives (at least until the final extinction program, which occurs after humans have ceded too much power to the AIs to stop it.) He could be right. Or not. In any case, the real world will likely be a lot more complex than his action-movie plot of a scenario.

“AI 2027” is a scary sci-fi projection

US Copyright Office affirms that only humans can be granted copyright

Last week the US Copyright Office released a highly-anticipated report on the copyrightability of AI productions, and affirmed that copyright requires that a human has had substantial input into the form of the finished work. Prompts alone will likely not pass muster, so some human-AI collaboration beyond prompting will be required. Meanwhile, federal circuit courts are hearing multiple cases where the boundaries of copyright in the age of AI are being explored. Probably most famous, and arguably most consequential, are the cases in which copyright holders, ranging from independent artists to the New York Times, are arguing that training AIs on their copyrighted works, even works freely available on the internet, are a violation. In response, the AI companies are simultaneously arguing vigorously in court that training AIs is permitted under “fair use” doctrines, while quietly cutting deals with large copyright holders to pay them an annual fee.

US Copyright Office affirms that only humans can be granted copyright

Brookhaven researcher solves physics conundrum with AI

Yesterday a researcher at the Brookhaven National Laboratory, one of the premier federally-funded research centers in the US, released a paper on an arcane area of physics which relates to how atoms are stacked in layered materials. These types of problems are so complicated that they are usually handled with approximation methods. However, in this paper, the author reveals how he used one of OpenAI’s top reasoning models, with the catchy name of “03-mini-high”, to arrive at an exact solution for one small subset of these problems. Not only is this a landmark result for this field, it is a landmark in using commercially available AI to assist with cutting-edge research at the frontiers of physics.

Brookhaven researcher solves physics conundrum with AI

MIT engineers develop flexible “vine” robot to search through rubble

Although humanoid robots get most of the press attention, work is proceeding on semi-autonomous robots in all sorts of body shapes. Last week MIT News ran an article on a flexible robot modeled on plant vines, made for rooting through the rubble after a disastrous collapse, in order to help rescue survivors. The robot, named SPROUT, can snake through small openings and bend around corners, carrying cameras and other sensors. SPROUT is already being field tested with Massachusetts first responders at their training facilities.

MIT engineers develop flexible “vine” robot for urban rescues

China licenses first 2 autonomous drone flying taxis for public use

China’s Civil Aviation Administration has licensed the world’s first two autonomous passenger drones, a major step toward full approval of commercial flying taxi services with unpiloted vehicles. Early uses for these drones are expected to include low-altitude tourism and urban sightseeing. The electric-powered 2-passenger air taxis have a cruising speed of 62 mph, a maximum height of just under 10,000 feet, and a 21 minute flight time, with recharging taking about 2 hours.

China licenses first 2 autonomous drone flying taxis

AI equals physicians in virtual Urgent Care diagnosis and management

Researchers from Cedars-Sinai Medical Center in Los Angeles implemented an AI chatbot for intake triage in a virtual, online Urgent Care clinic. The AI would make suggestions to the treating physician regarding possible diagnoses and care management for each patient. When comparing the AI’s initial assessment and plan of care to the physicians’ final diagnosis and care plan, the researchers found that the AI recommendations were of at least equal quality to the diagnosis and care plan decided upon by the treating physician. Findings suggested that the AI was better at identifying critical red flags, and supporting guideline-adherent care. Physicians were better at adapting their recommendations to changing information during the consultation. This raises the possibility that an AI-human collaboration in urgent care may be of significant value.

AI equals physicians in virtual Urgent Care

AI predicts adverse anesthesia events up to 10 minutes in advance

AI startup Predictheon, based in Barcelona, Spain, has developed an AI system that monitors a patient’s physiology during anesthesia. Trained with masses of historical clinical data, the system has learned to identify patterns that predict adverse anesthesia events up to 10 minutes in advance of their occurrence, allowing time for preventive interventions.

Predictheon’s AI predicts adverse anesthesia events