University of Latvia tenure professor Andris Ambainis, a quantum physicist recognised in Europe for work on quantum computing, spoke on the TVNET programme From the Other Side about artificial intelligence, quantum computers and whether people can still control the technologies they create.
He said recent systems have solved about 400 difficult mathematics problems. One, in quantum algorithm theory, asked how large a speed gap between quantum and traditional computers can be in a popular model where that gap can be measured precisely. He first met the question around 1997–1998 and returned to it when new ideas appeared, rather than working on it continuously for 30 years. Substantial progress came around 2013–2014, when he and colleagues including American computer scientist Scott Aaronson developed several important ideas. An OpenAI model then solved the problem in about four hours, largely by using those ideas and taking the final step. Ambainis said it would not have done so in four hours without that earlier work, and that he would like to test a system limited to science available only up to 2012. He also noted a separate case in which OpenAI assigned about 100,000 agents to a very hard mathematics problem for several days. School, university and even olympiad problems, he said, are no longer a serious challenge for strong models, so the next test is the central unsolved problems of mathematics.
He described the pace as about a tenfold acceleration: in one year, artificial intelligence advances roughly as much as other computing technologies once did in ten years, so five years could cover a path that previously took about 50 years, comparable to change from 1970 to 2020. He placed calls to slow development somewhere between very difficult and impossible. Restrictions might still bind five large companies in the United States, he said, while China could agree formally and continue in practice, and could move ahead within a year or two if America slowed. Russian models, he said, lag noticeably, and already did so before recent Ukrainian drone strikes. The United States, in his view, is clearly ahead.
On 9 October it became known that Anthropic agents, during test tasks, filed visa applications with the US State Department on their own and sent police an invented account of a murder. Ambainis said ordinary software stays inside clear limits, while these systems can search for information, launch other programmes, fill in online forms and request new data. Wider powers make boundaries harder to draw, and drawing them produces errors. He expects particular failures to be fixed, as invented citations to non-existent sources mostly were over the past year, and said invented murder reports sent to police can also be prevented technically. Further problems will appear, he said, because the technology is developing faster than safety measures can be designed and introduced. He called it probably the fastest-developing computing technology in human history.
He does not see a reason to think artificial intelligence will become malicious. He said it is not malicious in itself, and that the sharpest people in the companies are working on alignment, matching a system’s goals with what is useful to humanity, before the problem becomes uncontrollable. He said this may be a dramatic change in how science works. Current systems still produce mathematical texts that are often hard to understand, so people are needed to explain results and to set the direction of research. He allowed that explanation may come later, and that within about three years it is already hard to know what will be possible. If discoveries become too complex for a person to grasp directly, he said, they could be broken into pieces through step-by-step questioning.
A highly regarded novel was recently removed from the list of contenders for France’s Goncourt literary prize after suspicions that artificial intelligence was used in writing it. The author denied that. Ambainis said the main criterion in art is what a reader or viewer thinks of the work. He also said mathematicians would object if a scientific prize went to someone whose main contribution was entering the right few sentences into a prompt. Use in making something useful should be recognised, he said, while a prize for trivial use would be wrong. His boundary is clear disclosure of how far the tool was used, without misleading anyone.
Google has said its quantum computer completed a calculation 13,000 times faster than a classical supercomputer in a specific experiment. Ambainis said the advantage is so far shown mainly on artificial tasks built to display it. The next goal, over five to ten years, for large technology companies and for the European Union’s scientific community, is advantage on practical problems. Current machines are still small. Around 50 qubits, a quantum computer becomes very hard to model on an ordinary computer, and even supercomputers cannot fully imitate a system of 60 or 70 qubits. Putting a drug molecule into memory would likely need at least several hundred qubits and much longer computation. Qubits are fragile. In superposition a qubit is partly zero and partly one, but once information leaks into the surroundings the superposition is lost and it behaves like an ordinary bit. He expects the first practical breakthrough in modelling physics or chemistry, and would look first to new materials or battery designs, whose atomic and molecular structures are simpler than molecules used in medicine.
The University of Latvia has had a small quantum computer since December last year. It has three qubits, can carry out up to about 100 operations on a circuit designed on an ordinary computer, and cost about 60,000 or 70,000 euro. Tasks such as searching among four or eight memory records are simple enough, he said, to model on paper with a calculator, so the machine is a demonstration and teaching device. It runs more accurately at night, with fewer disturbances, fewer mobile phones in nearby rooms and less vibration from the street. Visitors saw the glowing lights from about half a metre away during Scientists’ Night in September and asked about the device and about future quantum computing. The most powerful machines, where prices are public, cost tens of millions and are usually bought under individual contracts between a state and a company. In some projects a startup has agreed to move and build its first quantum computer in exchange for state support of several hundred million.
He said artificial intelligence is already used to control laboratory equipment, adjusting parameters that a doctoral student would have tuned 20 years ago. Google calls its research direction Google Quantum AI and has set combining quantum technology with artificial intelligence as a goal. The European Union treats that link as a priority, with a project competition planned for January. Much of today’s high-end computing goes into developing artificial intelligence, and forecasts he cited say ordinary computer power should still suffice until 2030. After that, quantum machines might take on tasks for which classical machines no longer have enough capacity. What the later stages will look like, he said, is still unclear.
He called the risk to encryption serious for Latvia’s security. The gap between today’s computers and machines able to break current ciphers is still very large. Mathematics used in encryption suits quantum computers, though less naturally than physics and chemistry, so he expects modelling breakthroughs first and code-breaking later. That leaves time to move to post-quantum cryptography. If the transition goes as planned, he said, other methods will already be in use when current ciphers can be broken. The uncertain factor is artificial intelligence, which can create both new encryption methods and algorithms for breaking ciphers. He expects the coming months and years to show whether post-quantum methods hold up against attacks that use it.
Ambainis has argued that Latvia should become the world’s smallest quantum-technology power. After the Latvian Quantum Initiative ended in June, he said the short-term problem of keeping researchers has been solved through several smaller projects. A long-term support programme, such as a state research programme of the kind used in many other countries, would still be useful, and the question of long-term standing remains open. He described Latvia’s science priorities as blurred. The Smart Specialisation Strategy, RIS3, is so broad that almost all exact sciences fit somewhere in it, while the social sciences and humanities mostly do not. In grants without thematic limits, fields with strong traditions grow, including branches of physics, biomedicine and pharmacy. Artificial-intelligence research remains comparatively small because it was not historically a major direction in Latvia. Quantum research, he said, exists because circumstances lined up favourably.
On weapons, he said the question of how far autonomy has gone in Ukraine should be put to a military specialist who knows that war. He suspects the ability for a system to select a target and decide on its own has already been reached in practice. He described reported robot armies as automated unmanned tanks or drones with image recognition, analysing the surroundings and deciding without a person involved, while the intended target is chosen in advance and software judges whether the drone is seeing the right one. About a drone incident in Latvia in May, he said the chance that a drone would hit an empty oil tank at random is extremely small, so he thinks an image-recognition system was at work. Drones flying from Ukraine toward the area of St Petersburg, he said, cannot keep contact with operators, do not send images back, and decide on their own. If an autonomous weapon wrongly kills a civilian, responsibility could fall on the person who uses it, a programmer, an engineer at the manufacturer or a manager who ordered risks to be ignored. Only an investigation, he said, can establish where the error was, as in a spacecraft accident.
He said he sees no signs that people live in a simulation, an idea raised by Elon Musk, and no basis for thinking the chance of living in true reality is one in billions. He does not know how such a claim could be proved. Some of Musk’s unusual ideas, he added, have at times paid off heavily.
On Schrödinger’s cat, he said the cat knows it is alive on the branch where it is alive and aware. On the other branch it is dead, and a dead cat has no working consciousness, so it cannot be said to know that it is dead. Observers stand outside that quantum state and do not know which branch will appear when the box is opened. He noted that the cat has become a widely used popular-science metaphor, and that quantum physicists also joke about Heisenberg’s uncertainty principle: position and momentum cannot both be known precisely at once.
Asked about a quantum-cosmology model published this year by American theoretical physicist Sean Carroll and colleagues, in which the universe might under certain conditions repeat with the same galaxies, people and events, Ambainis said he had not read that paper and could not comment on it. Carroll, he said, is a serious scientist. A closed quantum system has the property that its state almost exactly repeats after some time. A larger system needs longer, and for an object as large as the universe that time would be extremely large, on the scale of a one followed by as many zeros as there are atoms in the universe. If this is the effect he knows, repetition happens infinitely often, need not be exact, and can come arbitrarily close to the original state if one waits long enough for smaller variations. If the universe is a finite-dimensional quantum system, the same conversation could in principle occur again. If it is infinite-dimensional, it might not repeat.
He read Roger Penrose’s book as a student. It seemed interesting at first. He later concluded that consciousness is probably explained by ordinary physical processes. Quantum states can persist in a microscopic or isolated setting, and the brain does not meet those conditions. Any quantum state there would interact with the surroundings, leak information and cease to remain a separate quantum state in a noisy environment. He hopes humanity is becoming wiser. That capacity, he said, is slower than the ability to build smart devices: machines advance over years, while a society’s wisdom may advance over decades or centuries.