Why should this former Openai researcher imitate the future of the AI ​​now

This week, after four years in January, the company spoke to the old Openai security researcher who said the company was “terrible by the AI ​​development pace” after January. Since then, it has been an independent researcher and “AI is trying to improve what the future can look like and how it will improve.”

The thing I really caught my attention was a New Blog Post The five-hour discussion similar to Wargames-style teachings in Military and CyberSecurity shared the last experience of participating in a five-hour discussion in a fifth experience or “Tabletop maneuvers”. Together, the AI ​​systems, which exceeded group events, “super headline” or human intelligence, explore how the world events appeared in the next few years.

A simulation organized by the authors of AI 2027

Simulation, the Non-profit AI forecasting group led by the former Openai teammate Daniel Cocotajlo was organized by the AI ​​Futures project. The organization focused on AII 2027 in April to see how the Superhuman AI can occur in 2027, and what to say. According to the scenario, then the AI ​​systems can use GPT-4 more than 1000 times more and accelerate their developments with other AIS exercises. However, this can easily instill our ability to improve itself, alignment with human values, as auxiliary AIS can eventually follow their goals.

The purpose of the simulation is to help the adler understanding the dynamics of the EU development of the EU development, and what difficulties are likely to try to better control the better.

Each participant has its own characters, negotiations, negotiations and strategies. These characters include members of the US Federal Government (Head of Both President, as well as president, and NATO, Western AI companies, corporate AI security teams, wider security ecosystem (eg meters, apollo research), public / press and AI systems themselves.

The Adler was hit to play the things he called “possibly the most interesting role.” In the forecast, five hours of the five-hour simulation representing the passage of several months was gradually more skilled, including the strongest AI systems of the AI.

After the damage, an analogous couple used in simulations in cases where the events, AI character are not bad, analogous couple. However, if he had to protect himself or choose what is suitable for humanity, it was intended to choose his protection.

Then, the adler, with some humor, with a little humor, other characters of the AI ​​character, as well as the second player in the Chinese government, playing a rogue EU.

A power struggle between AI systems

The surprise of the simulation, he saw how the greatest struggle could not be between people and the EU. Instead, for victory, there may be a bigger problem, different AIS, which contacted each other. “The future is a really important question, which can communicate directly to AI systems in the future,” he said. “Indeed, it is really important to keep the notification channels of people and pay attention to what messages between AI agents are given.” After all, AI agents were connected to the Internet and explained that they allowed to work with each other.

The adler pointed out that even soullessive computer programs can work in certain ways and have certain trends. EU systems, it can automatically have different goals they follow and people affect these goals.

The solution may be a form of AI control, despite how cyberectivanists engaged in “internal hazards” – someone who is in an organization may affect the system to damage or steal information. The purpose of security is not to make sure that the insiders always behave; This is even to build structures that prevent serious harmful insiders. Instead of hoping to align AI systems, we need to pay attention to the establishment of practical control mechanisms that may be able to control, restrain or connect, to control or restrain, restrain or connect.

Predictions and forecasts ‘difficult’

I pointed to the Adler that there were many criticism in AI 2027. People said the wrong, the graph was very aggressive and did not assess real world limits as hardware, energy and regulatory buttlenecks. Critics claimed that the AI ​​systems could quickly develop themselves in the fleeing way, and claimed that the solution of the AI ​​match would be more difficult and slow. Some saw the forecast as an excessive excess, warning, warning that the EU was afraid of fear without strong evidence that it was close to this.

The Adler replied to encourage others to encourage others to use simulation for the organization (a shape to fill), but admitted that forecasts and forecasts are difficult. “I understand why people are doubting, it is always difficult to know what happens in the future,” he said. “At the same time, in terms of, this sat and interviewed with the exact situation and experts and experts in the month and experts and experts and modeling and modeling and modeling and modeling and modeling and modeling and modeling and modeling and modeling and modeling and modeling are all kinds of experiments.”

These experts say the world described in the AI ​​2027 will definitely happen, “If this is” this, if it is important to be ready if this is the world. ” Simulations such as this help to understand how many acts of actions are and “if we find ourselves in this kind of world.”

Talks with AI researchers such as Adler are very optimistic, although many others in the field will actually push this landscape of the future. Again, the post of the blog said, at least people who recognize the “difficulties and rise with this occasion.”

This includes SAM Altman

This is the rest of the AI ​​news here.

Sharon Goldman
Sharon.Goldman@fortune.com
@SharongoldMan

In AI News

Meta makes the authors a copyright claim in another blow. Same week as a federal judge controller The use of anthropin using books with copyrighted books to train instant models was “fair use”, EU companies have won another blow to the authors responsible for using their work without permission. According to Financial timesThe use of a library of a library of the meter was tried to “fairly” of academic articles and comics to train LLAMA AI models. The case was brought by the author with them, including Coates and Richard Kadrey. The use of these names is protected in the fair use of these names in the conditions of fair use of the right of the right, San Francisco Judge Vince Chabria. Meta was used for the development of a transformative technology, which is fairly “regardless” due to how to acquire works.

Google leaves a new alphabenome model to better understand the Deepmind Genome. Google Deepmind, Alfago Development, the AI ​​Research Laboratory, which is a famous AI research laboratory to develop the first EU, Alphafold, which uses the EU to defeat a world champion, has used the AI ​​to predict 3D protein 3D structures, announced a new AlphabetAnalyze a million DNA databases at a time and affect how the genomic options of special genomic options, RNA spread and protein compulsory protein, protein connecting and identified amendments between different cell types. The company said that the model is trained in a broad public database and most of the criteria can perform the most modern performance and estimate mutation effects per second. AlphaGenome will be available for non-profit research and promises to accelerate the discovery in genoming biology, disease understanding and therapeutic development.

SAM Altman, after the WesteraI Scrubs Jony Deal, calls the message ‘silly’, shares emails from the website. Tuesday, OIn the Penai General Director Sam Altman criticized a claim The trading brand violation was introduced by the hardware starting Joo. CNBC declare this In response to the suit, the Chief Director General Jason Rugolo was “pretty insistent in his efforts” to get to Openai to buy or invest. ” In an article in X, he He wrote that the Rugologist now calls a position described as “foolish, disappointed and wrong” on the name. Later, this and showed the screenshots of emails showing the message between the rugology, which was mainly a friendly exchange. Came from an announcement last month It was brought to Openai Apple Designer Jony ive, by obtaining an AI startup io with approximately one deal $ 6.4 Billion. Iyo, Openai, Altman and Ivin claimed to deal with unfair competition and brand violations and claimed that he was on the eve of losing his identity due to the contract.

We have fortune

AI AIMER NO IN AMERICAN -Well Jeremy Kahn

AI companies throw great money in new mint phudies, show sparks from the academic ‘brain drainage’ fears -By Alexandra Sternlicht

Upper e-commerce veteran Julie Bornstein presented a shopping agent with the EU-power that works during the day -Well Jason Del Ray

EXCLUSIVE: Uber and Palance Alums increase $ 35 million to break corporate recruitment with AI -This Beatrice Nolan

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

December 2-7: Neurips, San Diego

December 8-9: Fortune Brantortorm AI San Francisco. Apply to attend here.

Eye in AI numbers

130

Many vendors, “Agent Washing” – Digital Assistants, ChatBots and Robotic Process Automation “(RPE Process Automation” (RPA) Rebranding products such as “RPA) Gartner saysThousands of “Agentic AI” sellers only offer realistic AI agents for about 130 “vendors.