Over the past few months, I've been using LLMs heavily at work. As a result, I have a newfound appreciation for the power of spoken conversation.
One reason is that it gets close to our true motivations. We speak what's really on our mind, and what our goals are.
If the conversation is recorded, a transcript provides a precious record which can help steer work using LLMs. A meeting transcript is one of the few text documents which can remain unquestionably human-authored and represents pure human insights (as long as people keep off their devices during the meeting!)
LLMs tempt us to go too fast for our own good. Things can get out of control and people get out of sync. Talking slows us down and helps keep people aligned on what needs to be done.
Implementing a solution is now faster than it used to be, compared to requirements gathering, ideation, and other planning. In my estimation, the calculus has changed in favour of more up-front planning. So the shape of a project looks a lot more like "lots of talking" followed by "a frenzy of agentic activity", even if the total amount of communication stays similar.
Agents are slow! Rather than spinning up multiple in parallel, which forces exhausting context switching and forces you to think about orchestration, we are often better to be having conversations while the agents are working. I have found that pairs-vibe-coding works well to this end: two people on the call, the reviewer providing oversight of someone using agents, and discussing the assumptions, approach, etc. This touches on another subject: deskilling and training new workers. I think LLMs help here too; let me explain why.
LLMs increase the level of competence needed to do my line of work, both since they automate away the simple stuff but also because they are a new tool in the toolbox, which means that the tool-user needs to be proficient in all the pre-existing tools, in addition to the new ones, not to mention the interactions between these tools.
LLMs are also a bit of a hazardous tool. People talk about AI psychosis, but I think context switching, the constant temptation to skip steps and circumvent review processes... these are also major hazards, even for someone who remains otherwise judicious. This means there's a higher level of skill needed to use an LLM.
Another aspect is that the commercial pressures are stronger. Everyone needs to innovate and keep up with the new developments to stay competitive, which makes the work more difficult than it used to be.
These all point in the same direction: the barrier to entry for professional computing is higher than ever. We can't expect educational institutions to produce workers who are immediately equipped for this new environment; we need more on-the-job-learning. More talking with junior staff helps move us closer to this apprenticeship model for training. Rather than expecting juniors to eschew LLMs entirely, or to give up on hiring them entirely, or something else entirely (as has been variously proposed) I think we just need to work more closely with them, and talk with them more. Help them see how we think. Help impart the skills to do whatever it is we're doing now. And, of course we're actually all still learning how this technology works and we all need to be learning from one another daily.
More talking means working more synchronously. Unlike many in my profession I much prefer in-person work environments. I prefer people to be back in the office.
Talking is a very human activity. LLMs are a de-humanizing technology. Nobody wants to read slop. Having a conversation cuts through the slop and forces people not to hide behind an LLM. I think subconsciously it probably makes us feel more accountable for our work if we have to speak to it regularly, which makes it more likely that people don't uncritically adopt LLM-generated work. As an aside, I think remote work is also somewhat de-humanizing.
More talking means more meetings, which is also likely to be unpopular. It's conventional to push back on pointless meetings, and I'm on board with that. So for these conversations, it's essential to prepare a clear agenda. For meetings where design decisions need to be made, that means clearly specifying the decision, the options, and the pros and cons of each. For important decisions, doing that thinking beforehand valuably forces people to slow down and leads to better outcomes than short-circuiting the decision by blindly delegating it to an LLM. For these sorts of alignment conversations, I'd recommend 2 or 3 people at most. Any more and some people don't participate (e.g. start getting distracted on their laptop, get dominated by the loudest voices, etc.)
I have found that meetings with a clear agenda can be timeboxed more effectively than threaded email or IM conversations. I think people take them more seriously and actually think about what they're doing! As a result, conversations seem more likely to get to the heart of the issue, whereas asynchronous conversation has a tendency to be superficial since people squeeze it in.
The need for a meeting can be proposed in a project backlog, with clear agenda attached. If people disagree about the need for a meeting with that agenda, then there definitely needs to be a meeting! If people don't agree that a decision is important to discuss, that's an important thing to resolve in its own right.
Talking fosters good communication skills. As we do less programming and more requirements gathering, etc., it means more stakeholder engagement and so more communication. So doing more talking helps us get better at that faster.
In a spoken conversation, we can gauge a lot more about how someone is feeling. We can help identify people's unspoken concerns. We can help get signs that someone might be feeling burned out. We can deliberately speak slower and draw contrast between the hectic pace of agentic programming and the slow and thoughtful zen that characterises successful project work holistically. We can better develop a sense of collaboration by trying to elicit opinions from voices who are quieter. In short, talking is a very powerful tool for setting (and monitoring) team culture in a way that asynchronous communication isn't.
With all this said, I also strongly believe in the power of written communication and of plain text. But I think that collaborative thinking is best done in conversation. With talking.