Once A Week

By W. F. Twyman, Jr.

[Note: Beginning on June 1, 2026, I will transition from writing daily essays to essays once a week. I enjoy daily writing but the demands and expectations I set for myself have been high. And I don’t like the feeling of falling short when I miss a day or two in a row. Weekly writing will allow more time for reflection and introspection. I have a novel, Gotterdammerung, to pull together for publication. My trusted writer friends have given me grace to observe life less, live life more. I love my writer friends like nobody’s business. As always, I thank you one and all for reading my observations on the human condition since March 23, 2023.]

AI Has the Power to Heal

As I type, AI is discovering remedies and cures for heretofore uncurable diseases. These miracles are occurring as scientists use autonomous AI to do scientific research. As a son of a mom who died from cancer in 1990, my heart rejoices upon hearing news of cancer breakthroughs. The discoveries below were never thought of before by human experts. The speed of AI research is that good.

For example, a paper in nature appeared ten days ago which struck me as a window into our immediate health care future. Nature is the most prestigious scientific journal in the world. The paper “Accelerating Scientific Discovery with Co-Scientist” by Google revealed the real world results of a system of specialized AI agents working together like a virtual lab. I am going to take a few moments and explain, as best I can, how this virtual team of autonomous AI agents works. I quote extensively from the you tube video below.

“Let’s suppose a human scientist has a research goal. The goal is to treat leukemina. The scientist assigns the ambitious goal to a supervisor autonomous agent. Let’s call this agent, supervisor. The supervisor takes the human goal and allocates the research to a team of underlying autonomous AI agents. The supervisor’s job is purely administrative. It’s job is to crack the whip and ensure underlying autonomous agents remain on task.

The supervisor assigns the research task to a generation agent. The generation agent’s function is to brain storm ideas, explore the scientific literature throughout the world, and synthesize findings into new proposed directions for a cure for leukemina.

Any output from the generation agent is directed to the reflection agent. A brutal reviewer of all ideas, the reflection agent’s job is to destroy hypothesis generated by the AI agent. Tear the ideas apart. Examine every proposed idea looking for flaws. Check for novelty.

The result is only high quality ideas survive.

The ideas now go to the proximity agent. Its job is to map all ideas together into a generated space. If the ideas are the same, the proximity agent says ‘stop wasting compute and resources and move on.’ Teams of AI agents can boost the speed of research. As a result, these agent teams have come up with new treatments for cancer, blindness, antibotic resistance and other diseases which are diffiult to treat.”

Want to know more?

*
”These breakthroughs are completely new. No human mind discovered these breakthroughs. Human experts never thought of these possibilities before. And we are not just talking theory. These advances actually work in real life.

The first of two papers as mentioned above is called Accelerating Scientific Discovery with Co-Scientist. Built by Google, this virtual team of AI agents is not a chat bot. Think of it as an ecosystem of AI agents working together.

In addition to the above discussed agents, there is also an evolution agent. This agent takes the generated ideas that survived a brutal review and refines them. For example, it might take two good ideas and mesh them together to produce one brilliant idea. The role of the evolution agent is to bridge logical gaps.

How does the system decide which of the surviving ideas is the best one? The system uses a ranking agent to host an actual automated tournament of ideas. The ranking agent uses an ELO ranking system. Just like in chess or competitive video games, if where a lower ranked player manages to beat a grandmaster, they get a huge chunk of their points. But if two equally ranked players tie, their scores barely move. It’s the exact mathematical principle, but applied to the rigorous evaluation of scientific ideas. And this co-scientist system actually simulates head-to-head debates between pairs of ideas. So, hypothesis A argues why its more new or plausible or testable than hypothesis B. And hypothesis B tries to argue back, pointing out flaws in hypothesis A. And then we have a separate AI model that acts as a judge, evaluating the debate.

Winners gain ELO points, losers lose them. And over hundreds or even thousands of these automated simulated debates, the strongest and most robust ideas naturally rise to the top of the leaderboard. Now, if you’ve playing around with AI, you’ll know that large language models are notorious for hallucinating sometimes just making stuff up. while sounding very confident. So, how do we know that co-scientist isn’t just making up these ideas? How do we actually know it’s producing better and legit scientific hypothesis? Well, the researchers gave co-scientist 15 incredibly challenging unsolved biomedical goals written by PhD level scientists and they asked it to come up with the best solutions. They then asked human experts to provide their asbsolute best guess solutions as well. They also gave the same goals to other state-of-the-art models at that time and they asked independent human experts to judge which ideas were better.

Now this is a blind test. So the human experts couldn’t see if the ideas were generated by other humans or by which AI models and the results were remarkable. You can see that co-scientists after giving it enough time to think and iterate eventually beat all the AI models and the human experts. These independent human judges rated the ideas from co-scientists as significantly higher in novelty, plausibility, and impact compared to the best solutions from human experts. So, this AI system, it’s not making stuff up. This is proof that it can actually create new innovative ideas which even independent human experts approve of.

Does this actually work in real life?”

*
The hour is late. I encourage you to review the video at your leisure for how the system was deployed in the real world.

The first example is acute myeloid leukemia (AML). This blood cancer is aggressive and, after a relapse, returns with a vengeance. The Leukemia stem cells are like weeds beyond the reach of conventional chemotherapy. What did the co-scientists do? They identified 2,300 drugs that were already FDA approved. Could one of these drugs hold the cure for AML? The prompt was simple: Which of these 2,300 existing drugs could be repurposed to specifically fight AML?

In the conventional world, designing a new drug from scratch could take a decade. Millions of dollars invested. A high failure rate.

Co-Scientist identified three drugs that could be repurposed to fight AML. These drugs were (1) Binimetinis, (2) Pacritinib, and (3) Cerivastatin. The scientists took the advice of Co-Scientist and tested Binimetinis, currently used for skin cancer, and tested it on AML cancer cells. It turned “out to be extremely effective.” A low amount of Binimetinis was found to be extremely potent against AML cancer cells.

The Co-Scientists decided to push the system. Identify a completely new drug for fighting AML that is unrelated to cancer use. Believe it or not, the system identified the drug KIRA6. No human had ever thought to apply KIRA6 to AML cancer cells. When the human scientists applied KIRA6 to AML cells, “The results were shocking. It wasn’t just good. It was remarkably effective.” 18 times more effective at killing leukemina cancer cells than standard drugs.

Human researchers had not thought of this method for treating AML cancer cells before.

Conclusion: Similar breakthoughs occurred with age-related mascular degeneration and antibotic resistance diseases. There is another automated system called Robin which solved a problem in less than 2 hours that would have taken human researchers around 400 hours to solve. And the price of the AI solution was $10.76. We are on the verge of significant medical discoveries in the coming weeks and months ahead.

These are good times to be alive.

Good evening!