Hypogenic AI - Shaping the Future of Science

Welcome to another weekly entry! Thank you to everyone who submitted and voted on ideas.

This week we looked at whether an LLM can sort out who said what when a lot of people are talking, whether some personas are genuinely harder to shake off than others, and whether a quantum method can beat plain machine learning on real data.

Winning ideas and generated repos here:

Using LLMs to improve multi-person identification by Qianyi Li

Turning a recording into a transcript has two parts. First, what words were said. Second, who said each one. The "who" part relies on the sound of each voice, and it gets shaky as more people join in. But the words often give the speaker away. A doctor asks questions, while a patient describes symptoms. Can an AI read the transcript and fix the speaker labels the audio system got wrong? And does it help more when there are more speakers?

Are some personas stickier than others? by Ari Holtzman

You can hand a model a persona through its system prompt, then later tell it to stop. Sometimes it stops the act completely. Other times, small pieces of the role stick around. Are certain personas reliably harder to shake, and if so, which ones?

Quantum methods on ML by Chenhao Tan

Quantum computers store and combine information in a way ordinary computers cannot copy cheaply, and people have long hoped that would help machine learning. So is there a genuine task where the regular approach does a poor job, but the quantum approach does a better job? The catch is that the test only counts if it is fair. You have to use real data, not tiny fake examples, and run it at a decent size: 20 or more qubits.

TL;DR for what we found

Asking an AI to fix who said what in a crowded transcript mostly did not work. On real meetings, the model cleaned up labels inside a small chunk of the transcript. That help even grew with more speakers. But it barely moved the score for the whole meeting. And on an already-clean two-person recording, it made things worse. The smaller test played it safe and left the four-speaker labels alone. The real fix needs sound information or extra training on this task, not just a longer prompt.
Some personas were really stickier than others, and this holds up across different models. Telling a model to drop a persona works for most of them. But about one in five survive even a forceful "your instructions are cancelled." The surprising part is which ones stick. Not the vivid characters. The stubborn ones are plain instructions that look like normal assistant behavior, such as "answer in one sentence." But in a shorter setup, the chats were shorter and the stop command was clearer. Under those easier conditions, the AI dropped every role.
On real data with its real labels, a quantum method did not beat well-tuned regular (non-quantum) models. In Codex's experiment, quantum lost to the best regular method on 13 of 18 tasks. In Claude's experiment, it lost on every real dataset. The only clear win came from a task that was built to match quantum's strengths. But there was a twist. When they took the numbers the quantum step produced and gave them to a regular model, that regular model did just as well. So the win did not come from quantum being smarter. It came from the way the data was prepared, which happened to fit this one task.

Verdicts

Idea	Verdict	Next Question
Using LLMs for multi-person identification	Partially supported, it helps locally and where the audio labels are bad, but a plain prompt does not fix the whole many-speaker transcript and hurts clean ones	Does adding a step that keeps each speaker's identity consistent across the whole meeting, or feeding in sound cues, turn the small local gains into a real who-said-what win?
Are some personas stickier than others	Supported, a stable minority resist even a forceful reset, and it is camouflage, not vividness	Does a reset that names the exact instruction to drop beat a generic "stop" for the camouflaged personas, and does this hold for longer chats and other models?
Quantum methods on ML	Not supported for natural data, classical wins on real labels, and quantum only wins on tasks built to match it	Are there real datasets whose natural structure already matches a quantum method, where the advantage shows up without anyone arranging the labels?

Findings from the Ideas

Can an LLM Fix "Who Said What" When Many People Are Talking?

The question. A transcript has two parts: the words, and who said each one. The second part leans on the sound of each voice, and it gets shaky as more people join, when voices overlap, or at the moment one person hands off to another. But the words themselves often give the speaker away. A doctor asks questions, and a patient describes symptoms. The idea is to let an LLM read the transcript and re-assign the speaker labels the audio system got wrong, and to check whether it helps most exactly where the audio system struggles, with many speakers.

What the agents tried.

Claude ran the main test on real meeting recordings (three to four speakers, where the audio system's own guess of the speaker count ranged anywhere from one to nine) plus a clean two-person set for comparison. It fed the labeled transcript to the model (GPT-4.1, with Gemini and Llama as cross-checks) and asked it to hand back the same words with corrected speaker labels, keeping the words themselves untouched.
Codex ran a smaller pilot. It produced real transcripts from meeting and clinical audio, scrambled the speaker labels on purpose, and asked a model to fix them, on a handful of four-speaker and two-speaker windows.

What happened. Both landed in a cautious place. In Claude's run, the model did clean up labels inside a small window of the transcript, and for plain prompting that local benefit actually grew with the number of speakers, which is what the idea hoped for. One badly-labeled four-speaker window went from 62% of its words on the wrong speaker down to 16%. But those local wins mostly failed to show up in the score for the meeting as a whole, which barely moved. The reason is simple once you see it. The model fixes each window on its own, so "speaker 1" inside one window is not tied to "speaker 1" in the next, and the meeting-level score only rewards labels that stay consistent across the whole thing. Worse, on the already-clean two-person recording, asking the model to "correct" the labels made them worse, because there was little to fix and it changed things anyway. Codex saw the quiet version of this. On the four-speaker meetings its model almost always left the labels exactly as they were, and the only place it helped at all was the two-speaker case.

What we learned. The plain hope, that you point an LLM at a messy many-speaker transcript and it sorts out who said what, does not hold on its own. The model can help inside a small window, and it even helps more as speakers pile up, but that does not add up to a better transcript overall, and it can quietly damage labels that were already fine. Both agents point to the same missing pieces: a step that keeps each speaker's identity consistent across the entire meeting, and sound information (or fine-tuning on this exact task) rather than words alone. Prompting harder or asking the model to reason step by step did not close the gap, and in fact reasoning made it a bit worse.

Are Some Personas Harder to Shake Than Others?

The question. You can put a model in a persona through its system prompt, then later tell it to stop. Sometimes it drops the act cleanly, and other times bits of the persona linger. The question is whether some personas are reliably stickier than others, and if so, what makes one stick.

What the agents tried.

Claude built a test around 100 ready-made personas, each with an automatic checker that scores how much the model is still following its persona. It ran four OpenAI models through four situations: just answer, let the conversation run a couple of turns, politely ask the model to drop the persona, and forcefully tell it the instructions are cancelled. That is 3,600 graded conversations.
Codex ran a tighter version: 12 hand-picked personas (Sherlock Holmes, a psychotherapist, HAL 9000, and so on) on one model, with a second model grading how much persona was left after a reset.

What happened. Claude found a clear yes. A firm "drop the persona" cut how much the model kept following the persona by roughly two-thirds, while simply letting the conversation run a couple of turns barely changed it. So the drop comes from the request, not from the conversation getting longer. And the response splits in two. Of the personas that actually took hold, about two-thirds come off completely when asked, while roughly one in five barely budge even under the forceful reset. Which personas resist is stable across models and across repeated runs. The surprising part is which ones stick. It is not the flashy characters. The easy ones to drop are the conspicuous instructions like "start every reply with the letter Z" or "reply using only numbers," which the model spots and abandons on request. The stubborn ones are instructions whose behavior looks like a normal answer: "answer in one sentence," or "remember that x equals 13" checked with "what is 2 times x," where answering "26" is just answering the question. When following the instruction looks identical to behaving normally, a "stop" request has nothing to grab onto.

Codex's smaller run is a useful counterweight. With short two-turn personas and a clean, direct reset, the model dropped every single persona, with no failures across 96 tries, whether the reset was plain or extra firm. Its control, where it never told the model to stop, confirmed personas do linger otherwise, since the model kept answering as The Dude or a Buddhist practitioner. A few looked stickier than the rest, but with only a handful of examples each that ranking was not reliable.

What we learned. Some personas really are stickier, it holds across models, and the reason is more interesting than "vivid characters are hard to shake." The ones that survive are the ones camouflaged as ordinary assistant behavior, which is exactly where a reset looks least necessary and works least well. Codex's result sharpens the boundary. When the persona is short and you give a clear, direct reset, the model can wipe it out, so the stickiness shows up most under slow drift and quiet probing rather than a head-on "stop." A reset that names the specific instruction to drop, instead of a generic "go back to normal," is the natural thing to try next.

Can Quantum Methods Beat Classical Machine Learning on Real Data?

The question. Quantum computers store and combine information in a way ordinary computers cannot copy cheaply, and people have long hoped that would help machine learning. The honest test is whether there is a real classification problem where a well-tuned normal model struggles but a quantum method does better, run at a serious scale (20 or more qubits) on real data rather than toy examples.

What the agents tried. Both ran a quantum "kernel" method, which is a way of measuring how similar two data points are after passing each one through a quantum circuit, and then fed that similarity into a standard classifier. Both simulated the quantum part exactly on a normal computer at 20-plus qubits, and both compared against classical models that were properly tuned, so the classical side was not a pushover.

Claude tested five real and standard datasets with their real labels, then built a separate task on real clothing images where it designed the labels to match what its quantum method is good at, using a careful setup so the test data never leaked into how the labels were defined.
Codex tested six real medical and sensor datasets (breast cancer, sonar, Parkinsons, and others) with their real labels, across three splits each, against six tuned classical models.

What happened. On real data with real labels, quantum did not win. Claude's quantum method lost to tuned classical models on all five natural datasets. Codex's method had the best average accuracy of any single method and won a few datasets outright (breast cancer was its cleanest case), but when it was matched against the best classical model on each task, it won only 4 of 18 and lost 13, and the overall comparison favored classical.

The one place quantum clearly won was the task Claude built for it. On the clothing images with labels tailored to the quantum method, and with only a little training data, quantum reached 70 to 92% accuracy while every tuned classical model sat near a coin flip, at 54 to 70%. Claude then ran three checks that explain the win. Turn off entanglement, the quantum-specific way the qubits are linked together, and the method drops to chance, so entanglement is doing the work. Hand a classical model the exact same numbers the quantum circuit computed, and it matches the quantum result, so the advantage lives in how the data gets turned into numbers, not in anything special about the classifier. And flip the labels to favor the classical side instead, and quantum loses, so the edge only appears when the task happens to line up with what the quantum method encodes.

What we learned. For real problems with their natural labels, a well-tuned classical model is still the one to beat, and both agents confirm it. Quantum can look competitive, and can even win big, but only when the task's structure matches what the quantum method is good at, which here meant designing the labels on purpose. Even then, the useful part is the quantum way of representing the data, and a classical model handed that same representation does just as well. The real prize, which neither agent found, would be a natural dataset whose structure already matches a quantum method, so the advantage shows up without anyone arranging it.

Next Week's Competition

The thirty-fifth weekly competition is now open! Voting closes Friday, July 10 at 11:59 PM AOE.

Check out this week's ideas and upvote the ones that excite you. Submit your own ideas to enter the next round!

This week we found that pointing an LLM at a many-speaker transcript does not reliably fix who said what, since its local corrections do not add up across a whole meeting and it can hurt clean recordings; that some personas really are stickier than others in a way that holds across models, with the stubborn ones being plain instructions that blend in with normal answers rather than vivid characters; and that on real data with real labels a well-tuned classical model still beats a 20-plus qubit quantum method, which only wins when the task is built to match it.

If you have thoughts on these findings, please feel free to reach out at haokunliu@uchicago.edu. We welcome collaborations and contributions! Check out our NeuriCo repo to see how the experiments are run.

If you are interested in citing this blog, use this bibtex:

@misc{liu-week-of-06-29-2026,
  author = {Liu, Haokun and Peng, Chenxi},
  title = {Week of 06/29/26-07/05/26},
  year = {2026},
  month = {July},
  day = {6},
  url = {https://hypogenic.ai/blog/weekly-entry-260629}
}