In the finish, the return took place after the reduction of 21 thousand. full -time jobs and the announcement of the “Efficiency Year” by Marek Zuckerberg. Joelle Pineau, leaving the position of the head of Fair in April, wrote: “When the AI race accelerates and the metal goes into a new chapter, you need to make space for others.” Her departure closed the era in which Fair developed far -long projects – From the optimization of drug therapy to self -regulating neural networks.
Genai, powered by billions of dollars, took over not only the GPU budget, but also people, and the priority was to introduce subsequent versions of the Llam and Meta AI assistant. Security? Science? Discoveries? All this already has a lower priority.
In Google, even the flagship Gemini 2.5 appeared without a full “model card”, i.e. the technical equivalent of food label. It was only a month after the premiere that the company supplemented the document with the results of the tests of dangerous capacity.
At the same time, Sergey Brin, returning to an active role in the giant, sent a note to Google Deepmind: “The competition has accelerated, the final race has started to Aga … We cannot still build overprotective products”. Memo directly ordered to shorten tests and work after 60 hours. weekly.
Openai, although still formally dominated by the Foundation, also cuts the procedures. The audit partner Metr admitted that he got a “relatively short time” for the assessment of the O3 model. Johannes Heidecke, head of OPENAI security systems, defended the decision: “We found a good balance between the pace and accuracy.” Researchers quickly noticed that while training AI models you need to take care of adequate vigilance because Later, you can only “powder symptoms”.
Check also: Each of us will be the head of AI employees. Change comparable to the arrival of the Internet
A great race to Aga
The most alarming is that the shortened procedures translate into tangible gaps. James White from Calypsoai summarizes the results of his tests in simple words. He says: “The models are getting better, but they are getting bad things easier.” This means Less repayments of the prompts describing, for example, the synthesis of biological weapons or infrastructure attacks.
A short idea – the product is forced not only by investors, but also physical limitations of infrastructure. According to IO Fund analysts Expenses of four giants for data centers will increase in 2025 to $ 330 billion, more than twice as much as in 2023. Each month of delay is hundreds of millions of dollars frozen in H100 chips.
Meanwhile, the Chinese Deepseek showed that a model with abilities similar to OpenAI O1 can be built, training it for $ 5-6 million, i.e. a fraction of the GPT-4 cost. Such a threat additionally motivates American companies to the prime minister in half the whistle, so as not to give the field to cheaper rivals.
Where are the regulations and, for example, procedures to protect copyright? These are just chasing reality, discussed dozens of times, often by people who fail to know the right knowledge and experience in the AI. The first bans of “systems with an unacceptable risk” from the EU AI ACT have been in force since February 2, 2025, but only next year the requirements of transparency for general purpose models will come into force.
In the USA, the White House enforces in turn Only voluntary obligations of companies to report security tests. In practice, this means that the company can shorten the test from six months to a week and still meet soft standards.
See also: Shorter full -time, the same performance. And will help companies balance a six -hour working day
Real costs for the market and users
Lowering the verification threshold of any AI model generates multi -layered risk. First of all, the exposure to Prompt-Injection attacks is growing-cases are increasingly reported when chatbots have made available API keys, customer data or instructions that should never go to a publicly available network.
Secondly, hallucinations become critical when the models go to search engines or medical applications. System online information reputation may suffer (He is already suffering) as much as during the appearance of the first fake news, but on an accelerated scale.
Equally dangerous are economic effects. If the success of the success is the speed of implementation, not reliability, Smaller research entities have no chance to compete with giants with the GPU fleet. This can lead to further market concentration and suppressing new research directions that do not promise immediate monetization. Paradoxically, the pursuit of Aga, justified by the productivity revolution, can slow down innovations outside of several corporations that will want to control the entire market.
If there is a serious incident-even providing a “recipe” for a biological weapon or a spectacular wave of Deep-Fake destabilizing elections-regulators will react, but almost certainly too late.
Is it possible to somehow regain balance? The Silicon Valley showed that he can transfer technology from the laboratory to a billion people during the year. We have evidence for this in success Chatgpt and Gemini. However, without a well -thought -out, independent of the research agenda profile, subsequent generations of systems can turn into black boxes with unknown internal goals. In the worst case, the economy that counts on a trillion of dollars of profits can discover that it has built a tool too unpredictable to use it safely.
In the long run, the market needs a hybrid model. Specifically, fast knowledge transfer to products, but also the requirement of public audit of key generations of AI models and transparent security cards. Only then the dispute between speed and caution will regain the proportionsand the race for profit will not be blind. For now, such changes are not expected.
Author: Grzegorz Kubera, Business Insider Polska journalist
