AI Opens New Horizons in Math and Life Science Research

The claim is that resolving the equation's blow-up problem would not alter actual phenomena like river flows or winds.

Conversely, some argue that mathematical proofs, like verifying the zeros of the Riemann hypothesis, hold intrinsic value separate from practical utility. Thus, solving this problem is not expected to directly threaten existing engineering algorithms.

This research employed a collaborative approach, operating 10,000 agents simultaneously.

Specifically, it applied the ansatz hypothesis-solving technique and cross-pollination, which merges ideas from various fields. Seungjun Choi reported that massive computing resources were expended, consuming 130 billion tokens over approximately 88 hours.

Seungjun Choi explained, "The model resumed post-training on August 28th. Then, from September 1st to yesterday, the results related to Navier–Stokes emerged from the intermediate checkpoints of that model."

Alpöge, a mathematician at Anthropic, has repeatedly stated on X (formerly Twitter) that he previously found breakthroughs in other mathematical problems using Claude and Claude's internal models. This led to speculation that OpenAI, aware of this, competitively invested massive computing resources into the Navier-Stokes problem.

Seungjun Choi said, "Alpöge had posted many times on X's timeline in August that he had achieved a breakthrough in another mathematical problem, not this issue, with Claude and presumably Claude's internal model."

At $50 per token, this translates to about $6.5 million, and likely over $10 million (approximately 13 billion Korean Won) in computing costs.

Noh Jeong-seok calculated, "If 130 billion tokens were used, at $50 per token, the output alone would be $6.5 million. And it must have been more than $50, right? Then $10 million is about 13 billion won." This significantly exceeds the $1 million prize for solving the problem, demonstrating the potential for overwhelming capital investment compared to human researchers.

There is ongoing debate about what substantial contribution these results make to the existing mathematical community.

Additionally, the temporal limitations required for Lean formalization verification and the absence of an official announcement from the Clay Mathematics Institute exacerbate the debate.

There are suggestions that the OpenAI model might have used researchers' hints as training data.

Seungjun Choi stated, "The scandal would be if the model somehow got hints from our inputs and used them." These suspicions have risen as a trust issue within the research community, escalating the controversy.

OpenAI researcher Jacob Coxson resigned, criticizing the irresponsible rush of AI development. Coxson expressed concerns about safety, stating that the company is gambling on superintelligence.

Evan Herbinger echoed Coxson's claims, mentioning the probability of AI posing a threat to humanity. They argued for the need for binding agreements beyond mere joint proposals.

Seungjun Choi said, "The Chinese response is that they can catch up. There are many posts saying that even though the frontier is now pushing ahead again, it's not a big problem, but rather that they are doing well and can keep up."

AI alignment work is underway, led by the Beijing government, and there is active movement in technological advancement, including rumors of the DeepSeek V5 release.

Like the metaphor in the novel The Three-Body Problem, there are concerns that if research results are revealed too early, entities with more capital and computing power can monopolize them.

Noh Jeong-seok said, "The moment coordinates are precisely marked, someone with more capital and more compute can come and take it." Consequently, researchers are re-evaluating their Terms of Service (ToS) and focusing on research security.

Bubeck et al. pointed out that Alpöge's research and OpenAI's approach differ in whether external forces were used.

Noh Jeong-seok stated, "OpenAI claims they handled it with external forces, so their approach is different. And they also said they gave a lot of autonomy to the model, but it turns out that humans still provided hints or guidance, leading to much debate." Contrary to the claim that the model solved the problem autonomously, there is ongoing dispute about human hints or guidance being involved.

Park Jong-hyun said, "There are many such reviews about Astra. You've probably used it a lot by now, and it's very good at computer use."

This means that for robot control, direct control is possible based on Vision, without a separate action model. Its applications in real-world tasks such as web analysis and robot control are rapidly increasing.

Park Jong-hyun stated, "Images 2.5 model has been released. One of the biggest challenges for image models right now is maintaining consistency."

He added, "This definitely maintains consistency much better. So editing works well. It seems they wanted to boast about this." This suggests the possibility of adopting an auto-regressive method similar to transformers, opening up potential applications in CG and VFX fields, such as stop-motion animation.

With the rapid advancement of AI technology, continuous follow-up is essential. Noh Jeong-seok said, "If you're not constantly following up on technology and you do it intermittently, then after three or four months, the world has progressed far beyond what you're using, and people seem puzzled by these processes."

Consequently, it is predicted that companies will launch agent package services, and the consumer market will expand.

Noh Jeong-seok explained, "I had Astra do everything. I ran it in Fast mode, and after High mode came out, I just surrendered everything and told it to book everything, and it finished in about 8 hours."

Astra handled everything from airline tickets and accommodation bookings to communication with guides using a 'Computer Use' method that directly manipulates the browser.

However, it showed limitations in payment approval processes, such as complex PG company authentication procedures, and the need for agent spending limits was raised.

Park Jong-hyun explained, "I asked Astra to create logos for our AI Frontier KR channel, but not just one. I asked it to create N options and then for Astra to directly choose based on its preference."

The agent generated four candidates and selected one itself.

The consistency of the generated logos and their applicability to merchandise like T-shirts and keycaps were also tested. However, some graphic deformations and errors in detailed elements were found.

The API cost for this task was approximately $2 for prompt-based work.

Noh Jeong-seok said, "Google announced AlphaGenome Atlas, so now you can simply look up all 9 billion mutations."

Similar to how AlphaFold revolutionized drug discovery by predicting protein 3D structures, this is expected to bring a major turning point in gene mutation research.

It has the potential to maximize the efficiency of designing disease target protein structures and drastically reduce the cost of genomic data analysis.

Research is underway on cell reprogramming using the Yamanaka factors (OSKM four genes) discovered by Professor Shinya Yamanaka. Noh Jeong-seok explained, "When proteins are expressed from these four OSKM genes, they trigger a mechanism that causes cells to lose their identity and revert to pluripotent stem cells, which can become anything, like when they first formed."

While these factors allow cells to regain youth by reverting to initial pluripotent stem cells, there's a risk of differentiation errors, specifically cancer. Recently, experiments with partial reprogramming using only three factors are being explored.

Noh Jeong-seok said, "Because there are risks, various switches are used to make sure these processes run in a very controlled environment."

They are exploring ways to enhance the safety of reprogramming, which rejuvenates cells in controlled environments, using OSKM factors and drug-activated methods.

HLA is located on chromosome 6 and is the most variable region.

Noh Jeong-seok explained, "If you read it with short-read sequences, you can't map it well. Because there are so many similar strands, the resolution of chromosome 6 is very low with short reads." In contrast, long-read sequencing reads genes in longer lengths, allowing for precise mapping of variations even on chromosome 6.

AlphaGenome differentiates itself from AlphaFold, which focuses on amino acid structure prediction. It provides the ability to predict experimental results for 11 cell types when 1 million gene sequences are input.

Noh Jeong-seok explained, "AlphaGenome is completely different. If you input 1 million gene sequences, it predicts 11 types of experimental results."

This is the result of pre-calculating and building a database for what happens if each of the 3 billion base pairs differs from the reference genome. This data-driven pre-calculation maximizes the efficiency of variant impact analysis.