Okay, take it easy. We’re not talking about the Singularity just yet.

The idea is that you can name your AI agents, to personalize them a bit, and make delegating tasks to them seem more natural. You can chat with them, too, verbally, thanks to newly engineered voice modes in models like GPT.

I was poking around and found this from Atomic Bot where the writer thinks up characters like “Concierge Clara,” illustrating the pros and cons of humanizing the bots. I like this, especially the second part:

“Keep the name to one or two syllables and check that it doesn’t mean something odd in another language.”

It would be awkward if you named your AI something that’s offensive to people in another part of the world.

You Can Give It a Profession, Too

Taking things a step further, you can also give the AI entity a job title. We’ll soon have AI doctors, AI teachers, maybe AI cab drivers, although that still makes some of us abundantly nervous.

One of the things I liked about a presentation I heard at the Next Endeavor event this past September in Mountain View, CA was the title: “AI Scientists are Here: Can We Trust Them?”

I thought this was one more instance of ushering in that age of AI vocation. But also: it’s about science, so let’s talk about some of that.

Prior to getting into a lot of the conceptual stuff, though, panelist Piotr Byrski of Molecule.one introduced a character named “Maria” who is, well, non-human.

“We built Maria,” he explained, “and Maria is a physical AI system that is capable of doing autonomous research in chemistry. So she comprises both the AI part, which is a bunch of agents and specialized models capable of the entire scientific workflow, from literature research through experiment design, all the way to data analysis, and a purpose-built lab where the experiments get executed and the feedback from the real world is provided.”

Here’s one of Maria’s notable achievements, given to me by Copilot:

“In a landmark 2026 collaboration with OpenAI, Maria AI and GPT‑5.4 optimized the Chan‑Lam coupling reaction — a key method for forming carbon‑nitrogen bonds in medicines. The AI identified primary sulfonamides as a challenging substrate class and proposed using mild oxidants like TEMPO. Maria Lab ran 10,080 micro‑liter reactions across two cycles, improving yields for 88% of boronic acids and 83% of sulfonamides. Mean yield rose from 16.6% to 25.2%, and the proportion of reactions above 30% yield increased from 15.6% to 37.5%.”

Give that lady a kewpie doll!

Panelist Scott Clark from Talaria Scientific mentioned some of the heavy stuff that scientists need to think about now, in the age of AI, regarding roles:

“I think that there’s this tension between two concepts. There’s the bitter lesson … which says that more data will just solve all problems. You don’t need to do feature engineering for natural language processing or computer vision anymore. But then there’s this also this concept of the ‘no free lunch’ theorem, that a specialized model that has better prior knowledge can beat a general purpose model.”

The group also discussed trade-offs with agentic systems.

“What you should be doing is trying to build AI that can adapt to the current lab,” Byrski said. “I think time has proven us right, that it should have been the other way around. It should have been (that) we are trying to adapt the way we do experiments to what we have achieved on the computational side, and not the other way around.”

“Can we use AI for that upfront design of experiments and then the translation to the robotics?” said panelist Nick Edwards of Potato. “Because typically, it’s the scientists who are doing this manually. They hand it off to an engineer who has to write up a script, to run it on the instrument. And so, really embedding these workflows into tools that agents, like you’re talking about, that are used, that can be used by humans now to help speed up their processes, fewer iterations, faster cycles, but then also enable them so that agents can be able to use them.”

“Fundamentally we saw a lot of these same trade-offs,” Clark added. “You have exploration versus exploitation. And it turns out many weak estimators can actually be better in that exploration phase to learn as much as you can about the underlying space and model. And then you exploit (it) with your very expensive systems.”

One of the other parts of this I liked was when Byrski talked about a term called “engineered serendipity.”

“The history of science is very much paved with luck,” he said, using penicillin as an example. “We can engineer ourselves to be lucky, because we can run so many experiments, we can analyze them, hence, engineered serendipity, we’re kind of poised to find something (because) there’s just no way that there’s no there’s no scientific knowledge resting in all that data, and that makes me very excited.”

And then there’s the speed of progress.

“I can do something in 6 hours that used to take me 6 months in my PhD,” Clark said, “which means I can ask more questions, I can read more papers, I can do things that otherwise would have been (difficult). But ultimately, by doing more of these experiments, by having that flywheel spin faster, we get a better surface area, and we end up climbing mountains that we didn’t even know we wanted to climb.”

At the end, moderator Frazer Anderson asked the panel “what’s missing from the stack.”

For Clark, it was more compute.

“A lot of the tools are there,” he said. “There’s a lot of great open source. It’s easier than anything ever to build things but we need more compute, and then more real world validation, which is what these two guys (other panelists) do.”

Edwards and Byrski wanted equipment.

“Drivers for equipment is a really big challenge in this space,” Edwards said, “because we have all this incredible scientific equipment that is hard to interface with. Anthropic has been doing some interesting things around that. I hope that there’s some convergence in terms of standards. But that’s a big one to me.”

“Equipment, equipment, equipment,” Byrski said. “Even those pieces of equipment that are, in theory, built for automated labs are, ultimately, in fact, built for human-led workflows. And so when you try to execute 100 times as many experiments on them, they just break down. So we’re actually speaking to a lot of vendors about this problem, to get, I think, (what is) a big opportunity … lab hardware is actually a hard market. I think it’s a big opportunity for that market to have a renaissance, and have a lot of growth in the coming years, if they figure this out.”

So that’s a review of the scientific role for AI, and a little bit about humanization, a kind of mixed bag, if you will, as we get closer to both Halloween and the Singularity.