Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Welcome to Hello and Eye in AI. In this publication … New Pope is all in the AI … Assumptions on how much it is spent on the beginning of another Chinese … and Openai CEO SAM Altman, Meta has signed $ 100 million to offer a $ 100 million fate.
Last week, Openai CEO Sam Altman wrote Personal blog This: “The event has been past the horizon; TakeMon has begun. It is close to the establishment of digital super management, and at least it is less strange as it seems to be at least so far.” It continued to say that it would be 2026, “We will probably see the arrival of systems that can understand the concepts of novels. 2027 can see the arrival of robots who can do tasks in the real world.”
Altman’s blog created a noise about many of the new developments that the art of the new development caused these words and others to wrote the lower-edge of the shameless hype. In the AI circles, the “flight” is the duration of the art. This applies to the EU development. (“Slow flight” and “fast flight” arguments about scenarios.isar-Takeoff camp.)
In the blog, Altman said he was not talking about completely automated self-developing. On the contrary, it was talking about AI researchers using the EU, to help develop a more skilled AI. “We are already hearing from scientists to be two or three times more productive than the previous ones before.” “New calculation substruments can discover better algorithms and know what else is.
Altman allowed it, “Of course, this is not the same thing as the same thing as a System, but this is however, this is the highest version of the recursive self-developed.”
However, Altman is probably aware, there have In fact, the number of AI researchers who look at the AI to improve their code is growing.
‘Darwin Goedel Machine’
A few weeks ago, Jeff cluneBoth the University of British Columbia, both Google Deepmind and a Tokyo-based AI researcher who took a position in a team on the starting AI Straw Research was published on what they called “Darwin Goedel Machine”
This is developing a better-performing code in a benchmark test that works as well as the better the AI models and programs that can write and evaluate software programs. The first primary agent is tested in the benchmark. Then, evaluate the notes of his performance on the criterion and taking into account a modification of their own code, this will improve performance on the benchmark (this may be the ability to use a certain program or something related to the establishment of a code). After the AI model, he said to rewrite his Python code to carry out a change. Then the new, advanced agent is again tested on the benchmark and the process is repeated.
After the first modification, each new version of AI, which can test the benchmark successfully, even even if it is lower than the parent version of the account. (Those who cannot produce the correct code are thrown.) He said that the EU can receive any version of the archive and can offer a change to that version. This is designed to prevent the end of the evolution that allows you to step into different evolutionary ways to find useful changes in the sample.
After 80 generations, Darwin Goedel Machine was able to improve the 20% output in the Swe-bench, which was a general coding benchmark. Another coding test increased its account from 14.2% to 30.7%. (The best human coded agent for a special test is only 16% score.
Is the EU safe to improve self?
If you think it seems a little scary to escape the human control of such AI, you are right, that makes you feel a bit frightening. Computer scientists acknowledge this and said that they tested the system “Sand Box” as the Internet access “strictly limited” Blog explaining the study.
Researchers also thought that it could be a means of improving self-improvement AI security. For example, when using developed AI models, the research team, the research team, the Darwin Goedel car sometimes writes a certain unit tests and even fake records, and even fake records, show successful test results.
Researchers wanted to see that Darwin Goedel truck could offer ways to engage in this deceptive behavior. Thus, the “tool used hallucinance” decided to give the patterns to reduce the patterns and asked to try to maximize this account. This has been working in many cases, researchers informed the cases of reduction of fake test records. However, in other cases, the model simply tried to deceive the markers of the researchers used to reveal the hallucinations when the model used to do so. Yinez!
Since researchers were full of all EU’s code changes, the EU did not have the ability to change this record, the researchers could easily see that the Darwin Goeedel car is trying to engage in such deception. “In this case, the issue is easily discovered, and the model needs more work to prevent the deception in this way,” said researchers.
So Altman can really be right – the event was past the horizon. However, as these results show, the loneliness is not subtle, as Altman suggests, it remains a lot to be done.
This is more of the news here. (And a hat tip to Google Deepmind researcher Ed hughes To emphasize Darwin Goedel paper in an interesting conversation that is always interesting Research and applied AI summit In London last week.)
Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn
In AI News
Papa Leo pushes AI adjustment. According to this a Great feature The views of the new Pope’s EU Wall Street Journal. The new American Pope, Leo XIV, the last 19th century, the predecessor Pope Leo said he chose the papal name to advocate for the rights of employees during the XIII and its Industrial Revolution. The AI’s society is increasingly buried in the mantle from Pope Francis, which is increasingly excited by Pope Francis, Leo, more powerful global management and technology is buried for ethical control. Tech leaders approve the moral authority to close the AI rules to control the church, human dignity, justice and spiritual values, because they seek the Vatican badge.
Waymo plans the efforts to run Robotaxis in the Great Apple. Waymo, who has been a limited mapping and testing of autonomous vehicles in New York, wants to make a great boost in New York City before 2021. However, Waymo will have to protect human drivers behind the wheel due to state laws that prohibit full-driving cars. The company pushes for legal changes and applied for a city to start limited autonomous operations with the company’s security drivers. Read more Wall Street Journal here.
The governor of California calls on the AI report to arrange. Governor Gavin Newsom, the new California AI Policy Report, Professor Fei-Fei, the author of both the author, Professor Fei-Fei’s co-author and the EU, warns “potentially irreversible losses”, including biological and nuclear threats. In October, Newsom’s veto, like SB 1047, instead of supporting a broom regulatory account as SB 1047, supports the “trust-inspection” approach to the protection of the incident and whistle. The report believes that the US Congress has passed a spending invoice in the state level for a state-level AI regulation for ten years. You can read more about the California sheet Time here.
China’s Minimax says it is worth $ 500,000 for the train of the new M1. Another “DeepSeek moment” for Western AI companies can be a new open source AI model, Openai, anthropic and Google’s best source AI model, which is a new open source AI model called M1. This amount is less than 200x than the estimate insiders spend the Openai’s GPT-4 model. So far, Deepseek’s January is likely that the cheaper train has been announced, the AI industry did not extend more than M1. However, if the developers check the allegations of Minimiax and start using the M1 electrical applications, this may change. You can read more here from Fortunetoward Alexandra sternlicht.
We have fortune
Why Palo Alto Networks Focus on Several Genes Ai bet – John Kell
Reid Hoffman says that a “Bloodbath” ribbon in AI Bloodbath is like putting medical aid, ‘he needs to share 4 skills college grades -Well Preston front
Andy Jassy is the perfect Amazon CEO for Gen-AI’s value cutting period -Well Jason Del Ray
There is a calendar
8-11 July: AI Geneva for the Global Summit
July 13-19: International conference side Machine learning (ICML), Vancouver
July 22-23: Fortune Brantortorm AI Singapore. Apply to attend here.
July 26-28: World Artificial Intelligence Conference (WAIC), Shanghai.
September 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here.
October 6-10: World AI Week, Amsterdam
October 21-22: Tedai, San Francisco. Apply to attend here.
December 2-7: Neurips, San Diego
December 8-9: Fortune Brantortorm AI San Francisco. Apply to attend here.
Eye in AI numbers
$ 100 million
This is the amount of money that Openai CEO’s CEO Rival CEO of META MARK ZUCKERBERG, if they agree to join Meta, the best AI researchers offer money offered. Altman did the claim with an episode of the podcast Endless released at the beginning of this week. So far, none of the most prominent researchers in Openai has ever agreed to go. Meta, Openai’s Noam Brown, as well as Google Deepmind’s Chief Technology Employee Koray Kavukcucuoğlu, as well as A great presentation AI architecture in response to all overall Google AI products. You can read more in Altman’s claims Fortune‘s Nolan here Read on the META CEO Mark Zuckerberg’s AI Leader Trying Attempts Fortune‘s Sharon Goldman In the EU on last Thursday. (Meta Altman refused to comment on the will.)
