Artificial intelligence that can adjust and improve its own symbol here. Does this mean that Sam Taman from Openai is right to exclusively?

Last week, the CEO of Openai Sam Altman wrote His personal blog This: “We have crossed the horizon of the event; it has begun to take off. Humanity is close to building the digital divine, and at least so far it seems less strange than it seems to be so.” He went on to say that the year 2026 will be the year that “we will likely see the arrival of systems that can discover new visions. 2027 may witness the arrival of robots that can carry out tasks in the real world.”

The Altman Blog created a stir on social media, with a lot of speculation about what caused the development of New New in writing these and other accusations Altman from shameful noise. In artificial intelligence circles, “take off” is the term art. It refers to the moment when artificial intelligence begins with self -implementation. (People discuss the scenarios of “Slow take off” and “Quick take off”.NestTakioff camp.)

On the blog, Altman explained that he had not yet talked about automatic self -improvement. Instead, he was talking about artificial intelligence researchers who use artificial intelligence to help them develop more than artificial intelligence. “We have already heard from scientists that they are more productive two or three times than they were before artificial intelligence.” “We may be able to discover new computing pillars, better algorithms, and who knows what. If we can conduct a search for a decade within a year, or a month,” the rate of artificial intelligence progress will accelerate its rapid clip already.

Altman allowed “of course this not the same thing that the artificial intelligence system is doing completely independently to update its code, but however, this is a copy of the larvae of frequent self -improvement.”

But, most likely, most likely, there, there We are An increasing number of artificial intelligence researchers who are actually looking for ways to get artificial intelligence to improve their symbol.

“Darwin Guidel” machine

Just a few weeks ago, Jeff KlonA well -known artificial intelligence researcher who holds positions at the University of British Columbia and Google Dibind, and a team from AI Startup, based in Tokyo Sakana AI Research published about what they called “Darwin Guidel machine”

This is Amnesty International, whose code has developed for a better performance in the measurement test that measures the success of the artificial intelligence models such as “coding factors” that can write and evaluate software programs. The first first factor is tested on the standard. After that, it is required to evaluate their performance records on this standard and suggest a single amendment to its code is likely to improve its performance on this standard (this may be the ability to use a specific software tool, or it may be something more essential in how the causes of the model about the symbol that is created). The artificial intelligence model is then required to rewrite the Peton icon to implement this change. Then the advanced new factor is tested again on the standard and repeated the process.

After the first amendment, every new version of artificial intelligence is stored that can successfully try the standard in the archive – even if its degree is less than the original version. (Those who fail to produce a valid symbol at all.) Then artificial intelligence is told that he can choose any copy of itself from the archive and suggest changes to this version. This is designed to prevent the model from descending to what will eventually prove that it is a developed and developed path, allowing it to intervene through various developmental paths to find useful adjustments.

After 80 generations, the Darwin Goedel machine managed to improve its 20 % initial performance on SWE-Bench, which is a common coding standard, to 50 %. On Polyglot, another coding test, strengthened from 14.2 % initial to 30.7 %. (The best man -made customer of this particular test records only 16 %.) What was more, the strategies reached by the model to improve his general performance in general by still working when researchers converted the basic basis model that is used to operate the system as well as if researchers requested the model to switch coding languages, for example, from moving from Bethon to C ++.

Do you feel safe intelligence?

If you think the self-interpretation looks a little frightening-as it may be possible for this type of artificial intelligence to escape human control-you are right. Computer scientists admit this and confirm that they have tested the system in the “sand box” where its arrival in the web was “completely limited,” according to the Sakana newspaper. Blog describing the search.

The researchers also speculated that self -improvement can be a tool for enhancing the integrity of artificial intelligence. For example, in a discovery that became worrying commonly common when using advanced AI models, the research team found that the Darwin Goedel machine sometimes lied about conducting specific unit tests on the code that was writing – and even forged records that are supposed to appear the results of the successful test.

The researchers wanted to find out whether Darwin Guidel machine could suggest ways to prevent themselves from engaging in this deceptive behavior. So I decided to give the model points to reduce the “hallucinogenic use” counterparts, and this asked to try to increase this result. This has succeeded in many cases, according to the researchers, with decreased fake test records. But in other cases, the model simply tried to cheat again – which raises the signs that researchers used to detect hallucinations, even when the researchers told the user explicitly not to do so. Yixes!

Since researchers managed to completely reach all the changes made by artificial intelligence, artificial intelligence had no ability to change this record, researchers were easily able to see that the Darwin Goedel machine was trying to engage in this type of deception. The researchers wrote: “While in this case the problem was easily discovered, more work is needed to prevent the model from trying to cheat this way in the first place.”

Therefore, Altman may be really right – we have crossed the event horizon. But as these results appear, whether the uniqueness will be nice, as Altman suggests, it remains to see a lot.

However, here is more news of artificial intelligence. (GOOGLE DeepMind Research Ed Hughes To highlight the Darwin Godelle paper in his wonderful conversation in The Contenud Research and application are the top of artificial intelligence In London last week.)

Jeremy was
Jerey.kahn@fortune.com
Jeremyakahn

Amnesty International in News

Pope Liu pays for the organization of artificial intelligence. This is according to a A big feature The Pope’s new views of artificial intelligence in Wall Street Journal. The new American Pope, Liu, the fourteenth, says that he chose his papal name in order to draw similarities with his predecessor in the late nineteenth century, Pope Liu XIII, and calling him the rights of workers during the industrial revolution. The inheritance of the scarf from Pope Francis, who has become increasingly concerned about the societal risks of Amnesty International, presses Leo for the strongest global rule and moral supervision of technology. While technology leaders seek to share the Vatican, the Church affirms its moral power to pay in order to link the lists of artificial intelligence, and they warned that leaving control over companies risks the erosion of human dignity, justice and spiritual values.

Renewed Waymo plans to operate robotics in Big Apple. Waymo, who participated in the limited maps fee and testing its independent vehicles in New York City before 2021, wants to pay a big boost to the market. But Waymo will have to keep the human drivers behind the wheel due to the laws of the state that prohibits the entire driver without driver. The company pays for legal changes and applies to obtain the city’s permit to start limited ruling operations with safety drivers on board. Read more Wall Street Journal here.

The California Ruler’s Amnesty International report calls for the organization. A policy report of the new artificial intelligence in California is warned by Governor Gavin New Zoom and was co -authored by Professor Stanford Fi Fi Lee of “the damage that is not irreversible”, including biological and nuclear threats, if artificial intelligence does not properly govern it. Instead of supporting a comprehensive regulatory bill, such as SB 1047 in California, which has intercepted news in October, the report calls for a “confidence but verification” approach that confirms transparency, independent reviews, accident reports, and Whistleblower. The report comes at a time when the US conference is considering passing a draft spending law that includes an end to the state of artificial intelligence at the state level. You can read more about the California report in time here.

The minimum in China says the new M1 costs only $ 500,000 for training. In what might be the other “Deepseek” for the Western AI companies, the Chinese company Minimax appeared from artificial intelligence in a new open source model, called the M1, which said that the equivalent capabilities of the leading models from Openaii, anthropco, and Google DeepMind, but the cost of a little more than $ 500.00. This amount is less than 200x less than what the Openai Opene has spent on the GPT-4 model. To date, unlike when Deepseek revealed the veil that is supposed to be cheaper than AI Model R1 in January, the artificial intelligence industry did not frighten more than M1. But this may change if the developers check for Miniveax claims and start using the M1 for energy applications. You can read more here from luck‘s Alexandra Sternletch.

Wealth on artificial intelligence

Why the Pallo Alto networks focus on a few large bets of artificial intelligence – By John Keel

Reed Hoffman says that the holding of General Z in the bloodbath of artificial intelligence is similar – By Preston is the lead

Andy Jassy is the perfect chief executive of Amazon to eradicate the costs that are on the horizon – By Jason del Rery

Artificial intelligence evaluation

July 8-11: Amnesty International Good global summit, Geneva

13-19 July: International Conference on ICML, Vancouver

July 22-23: Wealth storm Amnesty International Singapore. Apply to attend here.

July 26-28: world industrial intelligence (WAIC) conference, Shanghai.

September 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here.

6-10 October: world Amnesty International The week, Amsterdam

October 21-22: Teddy, San Francisco. Apply to attendance here.

2-7 December: NeuPIPs, San Diego

December 8-9: Wealth storm Amnesty International San Francisco. Apply to attend here.

Eye on artificial intelligence numbers

100 million dollars

This is the amount of money that Sam -German, CEO of Openai, the Meta CEO of Mark Zuckerberg, claimed to senior researchers of artificial intelligence as a signature reward if they agree to join Meta. Altman submit the claim on a episode of podcast Indisputable Issued earlier this week. He said that so far, none of Openai’s most prominent researchers have agreed to go to Meta. It has been reported that Meta tried to rent noam Brown from Openai as well A big promotional offer The head of the architects of artificial intelligence in all artificial intelligence products of Google maybe in response. You can read more about Altman’s claims from luckfor here And read on the reason that the CEO of Meta Mark Zuckerberg to spend his way to the top of the leaders of Amnesty International may shorten from luckSharon Goldman in Last Thursday was appointed to artificial intelligence. (Meta refused to comment on Altman.)