If you read the news about artificial intelligence, you may feel that you have been bombed with conflicting messages: artificial intelligence is prosperous. Amnesty International is a bubble. Current technologies and structures will continue in artificial intelligence in the production of breakthroughs. Artificial intelligence is on a non -sustainable road and needs radical new ideas. Artificial intelligence will take your work. Artificial intelligence is often useful to convert your family pictures into Ghibli Studio- Animation Pictures.
Confusion cut is 2025 Artificial Intelligence Index from Stanford the university The Institute of Artificial Intelligence centered around man. The pages of pages are 400+ are stuffed with graphic fees and data on the topics of research, development, technical performance, responsible spontaneous organization, economic influences, science, medicine, education and public opinion. like IEEE SICTRUM Do every year (See our coverage from 2021and 2022and 2023And 2024We have read the entire matter and expelled the graphs that we believe are telling the true story of Amnesty International now.
1. American companies abroad
While there are many different ways to measure the “presented” country in the artificial intelligence race (the articles of published magazines or citing them, Patents Grant, etc.), one direct scale is the one who sets important models. The Institute of Research has the era of artificial intelligence Database Important and important Artificial intelligence models This extends from 1950 to the present time, from which the AI Index drew the information shown in this scheme.
Last year, 40 prominent models came from USwhile China 15 and Europe 3 (by the way, all of this was from France). Another scheme, not shown here, indicates that almost all 2024 models came from industry instead of academic circles or government. As for the decline in the remarkable Models emitted from 2023 to 2024, the index indicates that it may be due The increasing complexity of technology and increased training costs.
2. Talk about training costs …
Yowe, but it is expensive! The AI index does not have accurate data, because many of the leading AI companies have stopped launching information about their training. But the researchers made partnership with Epoch AI to estimate the costs of at least some models based on the details that were collected about the training period, the type, amount of devices and the like. The most expensive model was to estimate the costs is the Gueini 1.0 Ultra from Google, at a picturesque cost of about 192 million US dollars. The general size of training costs with other results of the report coincided with: models also continue to expand the number of teachers, training time and the amount of training data.
It was not included in this scheme DibsicWhich rocked the financial markets in January with its demand to train a large competitive language model for only 6 million dollars – demanding that some industry experts They questioned. The coach, the guidance committee in the artificial intelligence index Yolanda Gille Recount IEEE SICTRUM It finds Deepseek “very impressive”, and notes that the history of computer science is filled with examples of early early techniques that give way to more elegant solutions. “I am not the only person I think there will be a more efficient version of LLMS at some point,” she says. “We did not know who would build it and how.”
3. However, the cost of using artificial intelligence decreases
The continuous increasing training costs (most) artificial intelligence models risk blocking some of the positive trends highlighted by the report: the costs of devices are low, the performance of the devices, and Energy efficiency What is up. This means INERFERE costs, or calculating the inquiry about a trainer model, decreases dramatically. This graph, which is located on the Logaretmite scale, shows the direction in terms of Performing artificial intelligence per dollar. The report indicates that the blue line represents a decrease from 20 dollars per million symbols to $ 0.07 per million symbols; The pink line shows a decrease from $ 15 to $ 0.12 in less than a year.
While energy efficiency is a positive direction, we will not return to negative: despite the gains in efficiency, total energy consumption has risen, which means that Data centers In the midst of an artificial intelligence boom, it has a huge carbon imprint. Approach the artificial intelligence index Carbon emissions One of the artificial intelligence selection forms on the basis of factors such as training devices, the cloud provider, and the site, and found that Carbon emissions increased from artificial intelligence models steadily over time – with an exotic dipceik.
The worst perpetrator of the perpetrator included in this scheme, Meta’s Lama 3.1, resulted in an estimated 8,930 tons of CO2 The emitted, which is equivalent to about 496 Americans living in a year of their American lives. This tremendous environmental impact explains the reason for the presence of artificial intelligence companies Nuclear embrace As a reliable source of carbon -free energy.
5. Performance gap is narrowed
The United States may still have a leadership progress on the amount of prominent models released, but Chinese models are of quality. This graph shows a narrow performance gap on the Chatbot standard. In January 2024, The US top model outperformed the best Chinese model by 9.26 percent; By February 2025, this gap narrowed to only 1.70 percent. The report found similar results on other standards related to thinking, mathematics and coding.
6. The last test for humanity
This year’s report highlights the undeniable fact that many of the criteria we use to measure the capabilities of “saturated” artificial intelligence systems – artificial intelligence systems get high degrees on the criteria that are no longer useful. This happened in many areas: general knowledge, logic about images, mathematics, coding, etc. Gil says she watched a surprise as a standard after the standard became unrelated. “I am still thinking [performance] It will go to the plateau, which will reach a point where we need new technologies or radically different structures “to continue making progress, but this was not.”
In light of this situation, the designer researchers have formulated new criteria who hope to challenge artificial intelligence systems. One of these is The last humanity examAnd that consists of very difficult questions that experts have contributed to the topic of 500 institutions all over the world. To date, it is still difficult even on the best artificial intelligence systems: the thinking model in Openai, O1, has the highest degree so far with 8.8 percent of the right answers. We’ll see how long it lasts.
7. A threat to the spread of data
today AI Tolide The systems get their intelligence through training huge amounts of broken data from InternetWhich leads to an idea often that “data is the new oil” for the artificial intelligence economy. As artificial intelligence companies continue to push the limits of the amount of data they can feed in their models, people began to worry about “peak data”, and when he ran out of things. One problem is that the websites An increasingly restricting robots From the crawling of their sites and avoiding their data (perhaps due to fears that artificial intelligence companies benefit from web site data with the killing of their business models at the same time). Web sites refer to these restrictions in Robots.txt files that can be read.
This graph shows that 48 percent of the data is one of the best web fields now fully banned. But a generation says it can end the new methods within artificial intelligence to rely on the huge Data groups. “I expect that at some point the amount of data will not be very important,” she says.
8. Here the company’s money comes
The corporate world has operated the Hanafi to finance artificial intelligence over the past five years. Although the total global investment in 2024 did not coincide with the mysterious highlands of 2021, it is worth noting that private investment was not above ever. Of the $ 150 billion in private investment in 2024, another plan in the index (not shown here) indicates that about 33 billion dollars went to investments in artificial intelligence.
9. Waiting for this return on the great investment
Companies are supposed to invest in artificial intelligence because they expect a great return on investment. This is the part in which people talk in countless colors about the transformative nature of Amnesty International and about unprecedented gains in productivity. But it is fair to say that companies have not yet witnessed a transformation that results in major savings or major profits. This graph, with the data derived from a McKinsey The survey shows that companies that reported costs, most of which were savings less than 10 percent. One of the companies that increased revenue due to artificial intelligence, most gains are less than 5 percent. This great reward may still come, and investment figures indicate that many companies are betting on. It is not here yet.
10. Dr. AI will see you soon, perhaps
Artificial intelligence of science and medicine is a small coating inside the mutation of artificial intelligence. The report lists a variety of new Basic models It was released to help researchers in areas such as Materialand WeatherAnd Quantum. Many companies are trying to convert the predictive and vulnerable forces into artificial intelligence The discovery of profitable drugs. and The OPNai thinking model recently recorded 96 percent on a standard called medqa, which contains questions from the medical council tests.
But in general, this seems to serve as another huge capabilities that have not yet been translated into a significant influence in the real world-perhaps, because humans still have not discovered how to use technology. This scheme shows the results of the 2024 study, which tested whether doctors will make more accurate diagnoses if they are used GPT-4 In addition to their model resources. They did not do that, nor did it make them faster. Meanwhile, GPT-4 surpassed itself on both human and human groups alone.
11. US policy actions are taken to the states
In the United States, this scheme shows that there was a lot of talking about artificial intelligence in Congress halls, and a little work. The report indicates that the procedure in the United States has moved to the state level, where 131 laws were transferred to law in 2024. DeepfakesProhibiting its use in elections Or to spread Uniform intimate images.
Outside the United States, Europe passed Artificial Intelligence LawAnd that puts new obligations on companies that make artificial intelligence systems that are high risks. But the great global trend was the countries that meet together to make comprehensive and non -binding statements about the role that artificial intelligence in the world should play. So there is a lot of talk everywhere.
12. Humans are optimistic
Whether you are a photographer, marketing manager or truck driver, there was a lot of public discourse about whether artificial intelligence will come in your job or when. But in a recent global survey about situations on artificial intelligence, the majority of people did not feel threatened by artificial intelligence. While 60 percent of the respondents from 32 countries believe that artificial intelligence will change how they take their jobs, only 36 percent are expected to be replaced. “I was really surprised,” says Jill. “Jill says,” Jill says. “It is very useful to think,” Amnesty International will change my job, but I still bring value. “Stay tuned to see if we all bring value by managing an eager teams of artificial intelligence staff.
From your site articles
Related articles about the web
