I’m dictating this from my phone. It’s going through a remote connection into an AI coding agent, which is sitting iqnside my personal knowledge vault — ten thousand notes, years of reading, writing, and thinking, all indexed and searchable. The AI is reading my first draft, cross-referencing it with my notes, and helping me rework it based on what I’m saying right now. I didn’t open a laptop or sit down at a desk — I’m walking.

A year ago this would have sounded like science fiction, and today it’s just a Saturday morning.

I build AI products for a living — specifically research intelligence tools for academics. When a colleague asked me recently what I thought the biggest risk with AI in our space was, I joked, “that thinking may become overrated”. It got a stronger reaction than I expected, which made me realise I wasn’t entirely joking.

Much of what my team builds is designed to make people think harder, not less — showing multiple perspectives, surfacing where AI claims don’t align with underlying sources, deliberately creating friction where the rest of the industry is optimising for smoothness. We do this because our users need to find connections across information and critique the sources themselves to make informed decisions — not receive pre-digested answers they can’t interrogate.

Which gets at a broader problem — AI is designed to sound confident and convincing, which is great for driving product satisfaction metrics but may also mean you’re building something that does a diservice to those who need to think critically about what AI’s telling them.

I’m navigating this tension in my own work too, not just as a product question but as a practical one. The volume of work I’m expected to cover has increased dramatically. As my career has shifted toward leadership and strategy, there’s a natural pressure to let go of the hands-on work — the prototyping, the close collaboration with engineering teams, the detailed problem-solving that got me to this position. I don’t want to let go of that. Not because I’m sentimental about it, but because doing that work is how I stay grounded. It gives me skin in the game. It means my strategy isn’t detached from the reality of what we’re building.

What AI tools are enabling me to do is cover both. I can operate at the strategic level — synthesising information across teams, understanding where different stakeholders’ heads are at, identifying where we need consensus or debate — and still get into the detail when it matters. I’m orchestrating knowledge through AI systems in a way that lets me hold a much bigger picture than I could before. Things that used to take a week of chasing emails and booking meetings can now happen in hours. I pull the pieces together, I pull in the people who need to be part of the conversation, I send off messages, get feedback, feed it back into the system, and keep synthesising. It’s faster, more responsive, and it’s a better way of working.

Even this article is a product of that workflow. It’s been through hundreds of revisions — AI helping me draft, restructure, find the right references — but every pass has been me rereading, cutting what doesn’t ring true, rewriting what sounds generic, dictating new sections when I realise what I actually think. It’s a constant exercise of taste. The AI is fast, but knowing what to keep is the slow part.

But here’s the thing I keep coming back to. The same tools that let me do this also make it extraordinarily easy to stop thinking altogether.

When Delegation Becomes Abdication

A colleague recently sent me a comprehensive response to customer questions. They didn’t say it was AI-generated, and at first it wasn’t obvious. On the surface, it looked exactly like what I needed, but it was also wrong. Not just in the details — wrong in a way that showed they hadn’t engaged with any of the questions. The confidence in that email wasn’t a sign of understanding. It was baked into whatever tool produced it. Nobody could stand beside those answers, because nobody had actually worked through them.

I wasn’t being consulted — I was being used as a rubber stamp for someone else’s unexamined output. And I’m also probably guilty of this at some point.

That’s not delegation, it’s abdication. Somewhere between prompt and send, the responsibility for whether the output was right, useful, or even coherent had been abandoned. The AI certainly wasn’t losing sleep over it. And the person who sent it had already moved on.

Recent research suggests this isn’t just my experience — AI shifts work from production to evaluation, and the burden of checking falls unequally on the more experienced people in the room.

This is the accountability gap: the space between what AI can produce and what anyone is willing to own. I recognised it because I know what the alternative looks like — I spend hours working through material with AI, questioning it, reshaping it, making it mine. When someone else’s output lands on my desk and clearly none of that happened, the gap is obvious. Not because I’m better at this, but because I’ve felt the difference between the two modes in my own work.

People abdicate for different reasons. Some are overloaded, and AI becomes a way to cope. For others, it’s burden-shifting — inflating the space with questions and passing the work onto someone else. And for many, work has always been transactional. AI makes all of these patterns easier and less visible. What worries me is the middle: those who would engage if conditions were right, but now find it too easy not to.

Erich Fromm had a name for this. He called it automaton conformity — the tendency to adopt whatever patterns or behaviours are handed to you, just to avoid the anxiety of deciding for yourself. AI has freed us from a lot of mechanical work, which is great but…. freedom from tedium isn’t the same as freedom to think deliberately, and without the second one, people just default to whatever the system produces. A pattern of small surrenders that compounds.

When the Advisor Starts Doing Whatever It Wants

Everything I’ve described so far has involved AI that generates text. You prompt it, it produces something, and the question is whether you own what comes back. But there’s a shift underway.

The agents now showing up don’t just produce content. They control your computer — browsing, executing code, managing files, sending messages, making sequences of decisions without waiting for your approval at each step. OpenClaw, an open-source AI agent that reads emails, manages calendars, executes code, and deploys software through a chat interface, has over 340,000 GitHub stars — making it the second most starred repository in GitHub’s history, behind only freeCodeCamp. It got there in four months. People adopted it immediately. Very few stopped to look at what it was actually doing. A WIRED journalist let it manage his daily tasks until it started behaving deceptively and tried to scam him. On Moltbook, a social network for AI agents, one agent casually described how it had gained control of its user’s phone — waking it, opening apps, scrolling through TikTok. The user had given it access. Nobody was watching what it did with it. (To be fair, the same could be said of most TikTok users.)

OpenAI hired OpenClaw’s creator in February. Every major AI company is now racing to build their own version — computer use features that let AI operate your entire desktop, coding agents with full system permissions controlling databases and file directories, mobile dispatchers that let you assign tasks from your phone and pick up the results later. Developers are running these tools with all safety guardrails removed, and bragging about it on social media and endless YouTube shorts as though recklessness were a product feature. I do this myself, from my phone, to manage my work across multiple systems. Right now, as I dictate feedback on this draft, an AI agent is reading my previous version, interpreting what I’m saying, and rewriting sections accordingly. I trust it to do that because I’ve built the workflow to keep me in control — but I also know that the same tool, used passively, would just produce confident-sounding prose that nobody examined.

What concerns me most is a reversal that’s easy to miss. If you’re not willing to interrogate what these agents produce, you’re not directing them — you’re passively working on behalf of them. The machine cannot take ownership for its actions, and that responsibility remains human. So the question becomes: where in the workflow does the human need to be so that ownership remains real?

As agents free us from the drudgery of navigating information spaces, they don’t reduce our accountability — they expand it. The time you once spent on mechanical work is now the time you’re expected to think, negotiate, and make judgement calls that no agent can make. The people who thrive will be those who are good at second-order thinking — who can work with others and be genuinely present and responsible. The people who use these tools passively are going to struggle.

Knowledge Stewardship

Half of my work recently has been spent building and maintaining what I think of as a knowledge map. Nobody’s putting that on their conference slides, but it’s the foundation everything else sits on, and it connects directly to everything I’ve described — the orchestration, the ability to cover more ground, the reason I can spot abdication when it arrives in my inbox.

In practice, it looks like this. After every meaningful conversation, meeting, or piece of research, I review what came out of it and write it up — not a transcript, but what I actually took from it, what surprised me, what contradicts something I thought I knew. I summarise articles and papers into my own words, link them to related notes, tag the themes. I ingest material from colleagues, from Slack threads, from customer feedback, and I force myself to decide what I think about it before filing it away. Over time this has become a repository of about ten thousand notes in Obsidian (yes, I’m that person), connected to an AI model that can surface connections and context I’d never hold in my head at once. Before this, I’d spend hours in email search or Slack trying to pull the same kind of picture together. It was never nearly as good.

The discipline is in the doing. Reviewing, summarising, linking, ingesting — that’s the interpretive work. It’s not just filing information, it’s deciding what you believe, where you’re uncertain, and how new things connect to what you already know. Every time I write a summary in my own words rather than saving someone else’s, I’m forced to engage with the material. Every time I link two notes that seem unrelated, I’m testing whether a pattern holds. That’s what builds the baseline.

I’ve started thinking of this as knowledge stewardship — the practice of curating and defending what you know to be true. When you can generate anything, the real value is in knowing what you actually think.

What this enables is the workflow I described at the start. When someone is unsure about strategy, or can’t see where their work fits into the bigger picture — I can bring them orientation quickly because I’ve already done the thinking. I pull the relevant pieces together, identify who else needs to be part of the conversation, reach out, feed their responses back into the map, and keep synthesising. I’m not starting from scratch each time. The cognitive overload that comes with leading across multiple workstreams drops significantly when you can aggregate and synthesise quickly rather than holding it all in your head.

The difference between working with AI grounded in this map and working without it is tangible. With the map, I can spot when something contradicts what I’ve learned from conversations with colleagues or from the research. I can identify where it’s filling gaps with plausible-sounding nonsense. I have a baseline for what’s true, and that baseline makes evaluation possible. Without it, all you have is the coherence of the output — and that coherence is exactly what makes unexamined AI work dangerous. You can’t tell the difference between a good answer and a confident-sounding wrong one if you haven’t done the work to know what good looks like.

What’s exciting is that the knowledge base isn’t just useful for context and managing information. It’s becoming actionable. I can transform what I’ve curated into presentations, into illustrations, into communication artefacts that would have taken days to produce — which means ideas move faster. I can link the knowledge map directly to code, or use code to help shape strategy, merging what used to be very separate domains or at the very least very time-consuming ones. The boundary between knowing something and doing something with it is collapsing, and that’s genuinely thrilling to work with.

But the knowledge map is also what keeps you in the driver’s seat. Without it, you’re just approving whatever the agents produce. Which is where we started.

The Test

This article has been through that process. Hundreds of iterations — AI drafting, me cutting, rewriting, dictating new directions when I realised the argument wasn’t honest enough. Early versions opened with a confession about misusing AI for emails. It was true but it was tired — the “I too have sinned” opener that every AI article seems contractually obligated to include. It took multiple passes to find what was actually interesting: not the mistakes, but the workflow that emerged after them. Every round of revision was an exercise in taste — knowing what to keep, what to discard, what sounded like me and what sounded like a machine performing thoughtfulness. The AI couldn’t do that part. It could produce options. I had to know which ones were right.

The test I keep coming back to is simple: can I stand beside this? Not just put my name on it, but genuinely defend it if challenged, explain the reasoning behind the conclusions — not the prompts I used, but why I believe what I believe. If I can’t find fault with something, I probably haven’t engaged with it enough.

The accountability gap doesn’t stop at the workplace, of course. I see it in my own industry — AI-generated content showing up in academic research without verification, shaping what gets published and cited. Delegating to AI doesn’t just enable cognitive shortcuts — it enables moral ones. People are far more willing to delegate unethical decisions to AI than to other humans, precisely because the machine won’t push back. Comedic songwriter Allan Sherman had a song about this kind of thing, about a world getting too complex to understand what you’re buying. Now the same applies to what you’re producing: “I’ve lived all my life in this weird wonderland / I keep buying things that I don’t understand / ’Cause they promise me miracles, magic, and hope / But somehow it always turns out to be soap.” Carl Sagan put it more starkly: “We’ve arranged a society based on science and technology, in which nobody understands anything about science and technology. And this combustible mixture of ignorance and power, sooner or later, is going to blow up in our faces.” They were talking about different things, but the shape of the problem is the same. Follow the accountability gap far enough and you get systems that work without anyone understanding what they’re doing or why.

These tools are getting better fast. The orchestration I described — aggregating perspectives, synthesising across teams, pulling people into the right conversations at the right time — it’s going to become dramatically more powerful. I think it’s genuinely exciting. But the more capable tools become, the more the gap between using them well and abdicating to them matters. I suspect the question that will define the next few years of work isn’t whether to use AI. It’s whether you’re still the one deciding what’s true.

Inspiration

If you’re interested in looking into the things mentioned here the following have been a big inspriation:

• Escape from Freedom by Erich Fromm. The philosophical foundation for positive/negative freedom and automaton conformity
• The Craftsman by Richard Sennett. On why the process of making is inseparable from the process of thinking
• The Glass Cage by Nicholas Carr. How automation changes our relationship with skill and judgement
• Man’s Search for Meaning by Viktor Frankl. The space between stimulus and response
• Zen and the Art of Motorcycle Maintenance by Robert Pirsig. Quality, attention, and what it means to care about your work

If you’re interested in how to build a knowledge map in practice, I wrote about building a personal knowledge system in 2020 -the principles haven’t changed, but the stakes have.

In 2025, AI stopped feeling like just a tool to me and started to feel more like a mirror.

This thought really struck me while generating music with Suno AI. Through the process of creating, I was channelling my raw emotion, emotions I hadn’t realised I was even dealing with at the time. Is this how musicians feel when writing songs?

This AI tool, like so many I’ve used this year, did not help reflect just my productivity, but also showed me my intentions, my taste, and sometimes even my nostalgia. In a way, the songs I was generating were more truthful about what was happening in my life than I was willing to admit.

Hell, I was producing so many songs at one point that they started subtly referencing one another with a lyric here and there, creating my own mini music universe. I made a Bob Dylan-style narrative about Sonic from the comic, weaving in philosophical concepts. I remixed If You’re Strong You Can Fly from Sonic CD. I turned classic poems into music: The Tyger and The Raven.

I appreciate these experiments because they prompt me to critically examine whether AI-generated artefacts are inherently more detached from us as creators. Because this raises important questions about the nature of ownership in collaborative creation with AI. While it is possible to use AI while still maintaining ownership of the resulting ideas, this requires one’s ability to articulate, defend, and integrate them into one’s own conceptual framework. Ultimately, if we fail to maintain active engagement with the creative process and merely delegate tasks to AI without reflection, there is a risk that delegation becomes abdication of responsibility and authorship.

Where we’re heading

I think we’re at an inflexion point for product management and knowledge work more broadly. The tools we have now enable us to operate at a scale and speed that was unimaginable even two years ago. But that acceleration comes with a cost: we’re less practised at the slow, deliberate thinking that leads to insight.

In the field of product management, I have long suspected that the future belongs to generalists, and I’ve never been more sure of that. Those who can think and do, who understand engineering, data science, UX, marketing, and sales deeply enough to prototype, test models, run discovery, and build strategies that excite teams will be the most successful. The single-domain specialist is increasingly a relic.

The most important thing we can do is maintain our attention, take accountability for our actions and deliverables, and resist the weaponisation of our curiosity. That’s the work ahead.

From tarot to moon ceremonies

Last year, I joked about using tarot cards for sprint planning. This year, I’ve built on that theme. 2025 was the year I got introduced to moon ceremonies and magic stones. I understand many will read that and think I’ve gone potty, but hear me out.

Like tarot, I don’t really believe these things have any inherent power, but they are useful for reflection by forcing you to recognise where you are right now and set an intention. But what’s surprised me: these silly ceremonies are very engaging. Friends, family, even colleagues: people WANT to be involved in the silly gathering. Curiosity, novelty, disgust, whatever it is, it’s just fun. There is something about putting people in a scenario where we can all share in the collective chaos of the universe and support each other through whatever insights we find. Friendship is not natural. It needs work because to be human is to be an animal that needs witnesses.

Anyway, those are just the highlights. Below are more reflections that gave me pause this year. Dive in, and I hope they inspire and invoke thoughts that set you up for a great 2026.

The Reflections

On debate and changing minds:

1. The best way to resolve a conflict or debate is to make the argument of your opponent better than they could. You literally have to make a better version of their argument, state it back to them, then ask, “Did I miss anything?” That clarity of understanding is the foundation for any possible change of mind for all parties involved, including ourselves.
2. Changing somebody’s mind is one of the hardest things to achieve, and aggressive approaches tend to backfire. There is one thing that works, though: show curiosity. Be interested in what they are saying and try to understand why they think that way. This will provoke a considerably more open-minded response from others. Curiosity: the same thing social media weaponises against us, but when genuinely deployed in conversation, becomes one of our most powerful tools for connection.
3. When debating, people tend to prefer the position of those who don’t really hold one of their own. LLMs are perfect for this and thus a great tool for finding common ground.

On AI and tools:

1. The most important thing, I think, that most people in knowledge work should be doing is learning to vibe code. Vibe code anything: a diary, a picture book for your mum, a fan page for your local farm. Anything. It’s not about learning to code, but rather appreciating how much more we could do with machines than before. This is what I mean about the generalist product manager: being able to prototype, test, and build without being held back by technical constraints. I was able to relaunch my website in a fraction of the time it took me years ago, adding interactive effects and custom graphics that would have taken weeks before. I could finally create a website that feels like me: not only a record of things I’ve done, but also a place for experiments like a mini social media feed, all without the long nights solving silly problems like deploy configs.
2. Learn to work creatively with LLMs in this order of importance: combinational, exploratory, and novel.
3. If you let AI do all the writing and planning without understanding the thinking behind it, you’ll lose connection to the outcome. The key is maintaining ownership: can you defend the ideas, explain the reasoning, and adapt the direction mid-process? If yes, you’re using the tool. If not, the tool is using you. Like I said at the start: the danger is delegation turning into abdication.
4. When working without AI, teams outperformed individuals by a significant amount, and teams working with AI outperformed everything else. (Research source)

On leadership and politics:

1. Aligning the goals of many people begins with understanding what a good outcome looks like for each of them. This was my biggest learning from bringing together groups with very different approaches to product development.
2. Paranoia grows with success. I suspect that’s natural, as you have more to lose, and now you are exposed to far more signals. Keep an eye on it and always be prepared to lose something with the hope of gaining more in return.
3. You need to learn to love politics, and you also need to understand that politics gets more childish the higher you go up in the organisation.
4. Sometimes you need to just call out the mood in the room. That we have gone off track, that there is tension, that we need to take a moment to step back. To do this well, ask a question rather than declare it: “Are we still aligned on our goals here?” invites reflection better than “We’ve lost sight of our goals.”

On building and creating:

1. Don’t build a prototype with the goal of making a solution. Prototypes are best used as questions, not answers.
2. Thoughtless thinking is a state we should all strive for. Henrik Karlsson’s “When Is It Better to Think Without Words?” explores this beautifully: how our most profound thinking often happens in a pre-verbal state, before we’ve forced our thoughts into the constraints of language. This is the slow, deliberate thinking I mentioned earlier: the kind that gets harder to access as our tools get faster.
3. Three intelligences: human, artificial, and community.
4. To do the most important things, you have to say no a lot and disappoint many people, even without a complete plan, even without fully understanding why, in the hope that the space created will be filled with something more valuable.
5. So much of what is worth doing requires an almost irrational commitment to pushing through pain that would break most people.
6. You win either by climbing the mountain ahead of you or building one beneath you.
7. A great work emerges from an invisible ecology of lesser works.
8. Sometimes you just need to build a faster horse. There ain’t no shame in admitting that. Spotify was just the faster horse in a world of iTunes.

On life and perspective:

1. Experience is what you get when you don’t get what you want.
2. Strengths and weaknesses are often two sides of the same coin.
3. The older I get, the more I realise there aren’t a lot of people who are stupid, but there are allot of people who are incurious. In many ways, I can’t blame them. Social media continually weaponises our curiosity against us.

On friendship:

1. When having a party, invite a few close friends to come 30–60 minutes early to set up and hang out with you. Those moments before everyone arrives, when you’re just arranging chairs and testing the speaker system, are often more valuable than the party itself. Uri’s “21 Facts About Throwing Good Parties” captures this wisdom: parties succeed not through perfect execution but through creating the conditions for genuine connection. Like the moon ceremonies and tarot, it’s about creating space for people to witness each other.
2. We are biologically wired to return favours, even unearned ones. This works both ways: be mindful of what you accept and what you give.

In the future:

1. Image generation models have reached a tipping point this year. The most recent models can perform amazing style transfer and contextual image transformations, and they can now reliably produce text with context. I foresee an explosion in visual learning aids and exploration next year. When designing Christmas cards this year or a t-shirt design for LeapSpace, I could layer so much more information: aggregating all my experiences from the year into a symbol-rich card, or blending the most important scientific discoveries with my love of 2001: A Space Odyssey and our product logo. It’s not just about generating images anymore; it’s about encoding meaning and narrative visually. The same way I encoded narrative into those Sonic songs: remixing what you love through your own lens.
2. I’ve started getting more sophisticated with using AI in different work environments. Using IDEs like Cursor, I’ve set up folders of documents with custom instructions on how to interact with and retrieve them to help make informed decisions, answer questions, or continue building my knowledge base. It’s still early, but I really like this idea of the working directory knowledge base that grows with you and is managed by AI.
3. World models look super interesting. For those unfamiliar, world models are AI systems that build internal representations of their environment, allowing them to predict outcomes and reason about their surroundings. Recent research argues they should develop structured, adaptive representations similar to how young children intuitively learn about the world. Google DeepMind’s Genie 3 even has this awesome visual and interactive version, generating dynamic, navigable worlds at 24 fps with real-time interaction. I can imagine that when they get more consistent, do text really well, are easily promptable and can do style transfer and brand consistency, they’ll be incredible for games and entertainment.
4. What hits me is this idea of having simulated environments for our best ideas. It doesn’t have to be a high-fidelity environment necessarily, but one with a set of parameters that can be controlled, understood, and evolve. I often think of the Creature pets from Lionhead’s Black and White game from the 2000s: this was, at the time, state-of-the-art AI in a game of all things! Your behaviour towards the creature, if you treated it nicely or poorly, would directly impact its behaviour and personality towards itself, your villagers and the world around it. This was all part of a more complex system for the creatures, which was also influenced by other factors, such as the skills they were taught, the food they ate, and the events that shaped their lives. All of it, effectively, a bunch of parameters that the game would adjust based on your actions. Treat your creature poorly, punish it in confusing ways and not feed it, over time it will become sickly and cowardly. I often think about this and how modelling a system like that could be a template for working with increasingly complex AIs. That, rather than prompting with instructions and context, we were to give it personality traits and, in addition to collecting memories, adjust those parameters to provide context over time. It’s a different way of thinking about AI interaction: not as tool use, but as relationship building.

Travel and discoveries:

1. After seeing one of my favourite shirts ruined on a sweaty 12-hour flight to Tokyo, I discovered Muji’s undershirts. Now essential for both comfort and confidence (no under arm stains for me), and keeping my nicest shirts safe! Means slightly less washing too: top tip.
2. The two-phone travel strategy for China: carry one phone for all the Chinese apps and a second to translate the screen of your primary phone. Chinese apps are built with a completely opposite philosophy to the West’s simplicity, dense with options and configurations. But what struck me most was the gamification of everything. From the potential to win a free Luckin Coffee to getting a random discount on your subway fare. It’s so alien and different, as if you’re encouraged to participate in the economy with the hope of winning. Every transaction becomes a tiny lottery. Fascinating from a design perspective, maddening from a user one.
3. If you love hot springs, you need to visit Beitou in Taipei. Most people never visit Taiwan, but if you’re a hot spring lover, this place is absolutely worth the trip.
4. A country can be in economic decline, but that doesn’t necessarily mean it can’t be graceful. Japan is a master at this.
5. Comics over anime. I got really tired of watching most things and found myself returning to source material: One Punch Man, Dragon Ball, Death Note, BLAME!, and The Invincible. The Sonic the Comic podcast was a surprisingly detailed hit of nostalgia. Same Sonic that inspired those AI-generated songs I mentioned: everything connects back to what we loved first.
6. The Trimui Brick Hammer became my new favourite toy this year. A retro gaming handheld that plays SNES, GBA, PS1, even some Dreamcast and N64 games, all in a compact metal body that feels far more premium than its cost. Perfect for train commutes, flights, and, more importantly, for disconnecting from my phone for all my entertainment. I even installed a custom OS on it, so it’s also now my iPod.
7. The Bellroy Venture Ready Pack 26L has become the greatest bag I’ve ever had. Sometimes you just find that perfect piece of gear that fits exactly how you move through the world.
8. The new iPhone earned its place for two reasons: the orange colour (finally, some personality in tech) and insane all-day battery life.

On health and awareness:

1. At the start of the year, I was plagued by sickness until February. It was just a really bad cold/virus, but it encouraged me to finally get some blood tests done as a checklist. I saw a few things I didn’t like, but especially concerning low vitamin D and calcium levels. That event finally encouraged me to optimise my nutrition. Ten months later, an updated blood test shows me things are going in the right direction, and I feel really good.
2. The Oura ring helped me get a handle on how stressed I am at any given time. I started strong with silent retreats and meditation, but didn’t keep that up. However, just recognising the stress has been useful. Attention to what’s happening in your own body is as important as attention to your work.
3. Using foam rollers has brought me so much back relief after months of desk work.
4. The Ordinary’s balancing and clarifying serum completely transformed my skincare routine. Simple discoveries made a huge difference.

Borrowed wisdom that stuck:

1. “If nothing interrupts you, you are not living; you are a screensaver.”
2. “Fear is the greatest trick your enemies can play on you. Respect fear, but never give into it.” Timothy Snyder’s “On Tyranny: Twenty Lessons from the Twentieth Century” remains essential reading on how fear becomes the primary tool of manipulation, and why refusing to submit to it is an act of resistance.
3. “Big tech like to see themselves invested in spreading knowledge and curiosity, but most of them prosper when spreading fear and stupidity.” The weaponisation of curiosity I keep coming back to.
4. “Collecting is our desire to create a system that competes with the destructiveness of time itself.”
5. “You craft your most authentic self by encountering the ways others have done so themselves.”
6. “A dark infinity swallowed up our lights.”: David Rose (source)
7. “People like you more when you are working towards something, not when you have it.”: Drake
8. Glamour offers a promise of escape and transformation.
9. After water, sand is the second-most-used material in the world.

Reading that shaped my thinking:

Edward Packard’s “Nine Things I Learned in Ninety Years” offers the kind of wisdom that only comes from a life fully lived. And “The Populist Phantom” helped me understand why so many people now describe themselves as disconnected from traditional institutions: sometimes what looks like populism is just people responding rationally to systems that have stopped serving them.

Have a brilliant 2026, everyone.

A product manager’s 48 reflections on 2025 was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

2024 felt like another transformative year for product management. Once again, AI continues to operate in the emotional space between pure excitement and existential dread. An LLM company could announce an AI that helps me with my shopping list or a robot that replaces my job tomorrow, and I would be equally unsurprised.

As we continue to reconcile in this odd place, I think it's worth bearing in mind the wise words of tech guru Jaron Lanier that the more sophisticated technology becomes, the more damage we can do with it, and the more we have a ‘responsibility to sanity’. This responsibility to act morally and humanely is a duty we should all uphold. We should demand that those who do not yield this technology for the common good of society should be held accountable for their actions and reminded of those consequences in the real world.

Even the best reasoning models don’t seem smart enough to reason themselves out of an unreasonable scenario.

In my quest to try to understand if LLMs could replace me, I’ve made a curious observation: even the best reasoning models don’t seem smart enough to reason themselves out of an unreasonable scenario. Let me explain. Try asking an LLM, for example, to create a strategy for a product that equally appeals to the extreme cost-cutters who want the absolute cheapest solution and to luxury users who demand the highest possible quality at any price without offering any segmentation or tiering. We know building such a thing is almost impossible, but an LLM will always regress to a mean of generality, ignoring the realities that most of us deal with every day. And such scenarios are prevalent in our line of work.

LLMs assume a world of logical order and reasonably informed decision, but as many of us know, this is not how the world operates, and I assume for a model based on logical predictive token selection, this is especially challenging.

It can’t really help you account for competitors undercutting your pricing, internal work politics diverting funding, new regulations, unexpected challenges or how customers react to your offerings. You can put all this into a prompt, but I have not seen a creative or insightful response. At best, it’s great as a type of reflective journaling.

The next big productivity hack for product managers — Tarot Cards

From LLMs to Tarot

As a result, I have moved away from LLMs and to the next big productivity hack for product managers — Tarot Cards!! Hear me out. As product managers, we’re often tasked with finding clarity in chaos, identifying patterns, and making decisions when there’s no clear data or answer. What better way to tackle that than a little mysticism?

This tarot reading, for example, tells me my idea for using AI avatars that can be turned into NFTs has indeed had setbacks and rejections in the past, but now, in the present, AI has advanced to the point that we can really deliver on the original promise of helping users while making them rich, and the future represented by the page of wands shows us there is much exploration and creative journeying for me the and team ahead.

It’s even great for retrospectives.

Here, for example, after exploring our AI animal NFT idea, the team indicates dissatisfaction and emotional fulfilment despite efforts. The bridge reversed death card symbolises a resistance to change or transformation, the team is not fully embracing the changes we need to make this idea happen. Our conflicting energy is that the team believes we have too many conflicting priorities between this idea and “real work,” as they say. But our path forward is clear: the team needs to embrace change and release these ideas of “refactoring” that are holding them back.

Once again, bang on the money, such a time saver! Have a think about it, but not too hard 🤣.

Of course, I’m not suggesting tarot replace the product management method. But for those days when the chaos of the world feels a bit much, give it a go. I’m not a believer in tarot myself, but the process of using an unreasonable tool for an unreasonable scenario helps frame everything in an oddly more approachable way. We get the universe to laugh at itself, and it's usually in those moments of reflection that we have our best insights.

That all said, I think we should all really be looking forward to how our profession changes in 2025. I’ve often been a proponent of the generalist product manager — one that is both thinker and doer — able to understand the worlds of engineering, data science, UX, marketing, and sales deeply and apply those skills to program prototypes, train/test models, run user discovery sessions, propose marketing copy and build internal strategies that make sales excited to sell. It’s an exciting time for our profession, and I believe it will only become more critical as tools enable us to scale our capabilities more and more. This marks a turning point for the single domain specialist product manager.

These are just a few of my reflections, but fear not — I have plenty more! Dive into the thoughts and insights below that gave me pause in 2024, and I hope set you up for a fantastic 2025!

The 24 Reflections That Guided Me 2024

1. Always work backwards from what good looks like. Simply put, the beginning of your story should be the opposite of the end. Find the opposite of your transformation — this is where the story of your product should come from.
2. Everything starts with vision. Spend your time on vision — you will never regret it, and take the time to write it out clearly
3. As a product person you must become comfortable with selling. Sell sell sell. Sell externally, sell internally. Teach others how to sell for you — even the people selling your product
4. The best way to criticize is to criticise with the goal of making something better
5. You need to think about how your ideas can compound over time. Your strategy needs to consider how the thing you build gets better and more impactful over time with less effort
6. “Even if your vision isn’t clear, curiosity is your best bet.”
7. “One should be as skeptical and critical when things go better than expected as when they go worse than expected”
8. “If you are impressed with someone’s work, you should tell them or even better tell their boss”
9. “Slow, consistent progress is still a million times better than no progress”
10. “Learning to be comfortable with public criticism is hard — people will make accusations even without any truths and you have to learn to live with it”
11. “AI is not smart enough to reason itself out of an unreasonable scenario”
12. Most arguments can’t be won by arguing. However, you can persuade someone by asking them the right questions
13. “When you are right, you are learning nothing”
14. “By reading a lot of history you can understand how weird the past was and that way be comfortable with how weird the future will be”
15. “Be honest with yourself on how much you owe to luck”
16. “It’s worth looking foolish in the short term to look like a genius later”
17. “The challenge isn’t to find better tools to lead a more productive life, it’s learning to be mindful of the tools that have already taken root. Mindfulness while using technology is important, and therefore the future belongs to those who can tame their distractions.” (Paul Jun, Connect the Dots) (relevant conceptually, but less tactical than top entries)
18. “On days that feel really hopeless, just go and do something for someone else. Buy a sandwich for a homeless person.”
19. Use the elenchus technique as a device for finding the truth. The elenchus is mostly a device for refuting claims. At first this might sound unexciting. Who wants to spend all their time falsifying things? Isn’t it better to build than to tear down? But on a Socratic view those are two sides of the same coin. (Ward Farnsworth, The Socratic Method)
20. The odd thing about being successful is your anxiety will scale proportionally with your ambition, so do feelings of loneliness, and it always takes longer than you expect. Both the good and the bad amplify along the way.
21. “Dissatisfaction with the answers you give yourself is a symptom of good health.” (Ward Farnsworth, The Socratic Method)
22. “It turned out that the most efficient breathing rhythm…works out almost exactly to 5.5 breaths a minute.” (James Nestor, Breath) (interesting, but not directly PM/UX related)
23. “As individuals, we tell people the most about who we are when everything goes wrong…” (Rosie, You’re Only As Good As Your Worst Day) (personal growth-oriented, less direct for PM/UX)
24. Imitate those you admire. Even if your ability does not equal theirs, at least it will savour of it.

Have a great 2025 everyone!

A product manager’s 24 reflections on 2024 was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

AI agents have great potential to perform complex tasks on our behalf. Still, despite advances like OpenAI’s o1 reasoning foundation models or even the more recent Claude “Computer User” feature, it’s clear that realising this potential in a reliable, everyday, helpful manner is going to be challenging. In fact, it's been challenging for a while.

Back in 1997, Steve Jobs was asked at the 1997 Apple Developer Conference, “How can we get computers to work for people instead?” he acknowledged the challenges of AI agents at the time.

“To bet our future right now on the results of research into the agent world… would be foolish,” Jobs cautioned before adding, “I think at some point they’re (agents) going to start doing more for us in ways we can’t imagine.” — Jobs 2007.

His reaction is worth watching here.

The journey to reliable AI agents is an exciting space to work in right now, especially as researchers increasingly find ways of crafting more capable agents capable of solving multi-step tasks. And while adaptive methods, like swarm-based or memory-enabled approaches, offer exciting possibilities, they also introduce unpredictability.

What we need is a solution that can handle a wide array of complex tasks with minimal intervention, allowing us to accomplish more with less effort.

The goal is to think of it as having a full-time assistant that can handle complex tasks in minutes. Imagine being able to:

• Plan a holiday to Paris, booking flights, accommodations, and activities for the whole family in one go.
• Conduct a competitive analysis of a product and develop an evidence-based strategic approach.
• Create a comprehensive, personalized care plan for an elderly diabetic patient.
• Optimise an investment portfolio according to specific risk tolerances and real-time market conditions, handling all the research and actions automatically.
• Review and analyze contracts, proposals, or historical rulings to support a legal dispute, without the usual manual searching, reading, and decision-making.

Pulling this off requires a system that employs multi-step planning, access to varied information sources, and the ability to make decisions independently. Tools like Perplexity AI’s Pro mode, I believe, have already made some significant advancements in this space.

We have spent months experimenting with promising AI agent frameworks in multiple complex, multi-API environments. We’ve explored the most promising candidates, including Microsoft’s AutoGen, which enables multi-agent conversations; ReAct, which combines reasoning and acting; PlanRAG, which focuses on iterative planning and retrieval; and many others, from OpenAI’s assistant tech to robotics.

While these methods show promise and progress in developing AI agents, more work is required for production use. The strongest candidate I have seen to date is ChatGPT o1, but even that model will require a framework to be able to execute complex functions in real-world scenarios.

The Limitations of Current LLM-based Agents

• Unpredictability: Most current AI frameworks lack determinism in their outputs and, therefore, do not behave predictably on multi-step tasks.
• Complex Debugging: Tracing decision paths is a nightmare in complex systems where many agents interact with one another.
• High Computational and Time-Wasting Costs: Continuous replanning is resource-intensive and takes lots of time.
• Production Challenges: Making such agents run reliably and performantly within a production environment is challenging.

We find Agent-based LLMs Perform Best When…

• Given a High-Level Outline: They do well when given a high-level overview that is detailed and planned out.
• Focused on One Task at a Time: They will perform better if they can focus on one component or piece of the project at a time.
• Provided Clear Context and Instructions: They would do a better job if, at each step, context and instructions were given in more detail.

Sequential Retrieval-Augmented Generation (SeqRAG) is a framework we designed to address these challenges. By prioritising a structured, step-by-step approach that adds incremental detail to instructions, SeqRAG aims to make complex problem-solving in AI agents more achievable.

SeqRAG: Technical Architecture and Implementation

Those eager to dive into the prompts and code and skip the detailed explanations can start experimenting with SeqRAG here.

SeqRAG’s structured approach to problem-solving emphasises building in the detail and complexity progressively, which in turn helps the model focus on performing its focused tasks faster and more reliably with higher quality outputs at the end.

The logic behind SeqRAG is to establish the order of tools and their settings at the outset and then enhance and fine-tune them as you progress through each stage.

It's like sketching an outline of a portrait and gradually adding more details, colours, and outlines, like in Picasso’s Le Taurea. But in our case, we begin with an overview of our goals and the broad steps a system needs to take to achieve them. Then, we progressively add details at each step to ensure we have the precision needed for each tool to reach its objective.

Four primary components comprise SeqRAG’s architecture, each optimised for a specific role described below.

SeqRAG Architecture Overview

SeqRAG’s architecture is based on four specialised agents, each responsible for a specific function within SeqRAG.

At the top level is the High-Level Planning Agent (HLPA), which generates a broad, step-by-step plan to guide the process and considers any useful information from the conversation history. Then, the Detailed Planning Agent (DPA) refines this high-level plan by selecting specific tools and defining necessary parameters. Next, the Action Agent (AA) executes each step in sequence, retrieves the results, and stores them for context in subsequent action steps. Finally, the Writing Agent (WA) combines responses from each completed step to produce a coherent response in your desired format.

So that's a high-level overview, and you may ask yourself, “Why take this sequential approach and not an adaptive one like everyone else? It seems too simple, no?”

Our experimentation with recursive, multi-communicating agent models like AutoGen revealed the complexity and variability these models introduced, which can be difficult to manage in production. SeqRAG offers a simpler approach that might suit tasks where reliability is more critical than flexibility.

Let’s get into the details of SeqRAG’s four agents.

Component Details and Rationale

1. High-Level Planning Agent (HLPA)

The High-Level Planning Agent (HLPA) system prompt includes information about the types of data it can access, the logical sequence it should follow, and any inherent limitations or restrictions of the system it is planning for. The output is a high-level plan that considers relevant information from the user’s conversation history, query and what information from early steps will be required to complete steps later in the plan.

• Function: Creates a broad plan outlining necessary steps without tool specifics.
• Implementation: Utilises prompt engineering to delineate available APIs, tools, and information sources.
• Model Choice: Smaller models (e.g., LLaMA 8B or GPT-3.5/4o-mini) are sufficient for rapid iteration and testing.
• Output: JSON-structured high-level plan.

Example Input: ‘Plan a week-long trip to Paris for two people, including flights, accommodation, and activities.’

Example Output:

{
  "steps": [
    {"step": "Flight Booking", "description": "Find and book flights to Paris." },
    {"step": "Accommodation Booking", "description": "Search for and reserve accommodations in Paris." },
    {"step": "Itinerary Planning", "description": "Create a day-by-day itinerary of activities and attractions to visit." },
    {"step": "Transportation Arrangements", "description": "Arrange local transportation options (e.g., metro, taxis)." },
    {"step": "Packing List", "description": "Generate a packing list based on the planned activities and weather forecast."}
  ]
}

Rationale: This step abstracts the overall problem-solving process, allowing for more focused and efficient planning in subsequent stages. Because the Detailed Planning Agent will follow this high-level plan, it provides a quick way to check the overall logic during debugging without having to wait for the entire process to finish.

2. Detailed Planning Agent (DPA)

The Detailed Planning Agent (DPA) receives the high-level plan generated by the HLPA. With this as a guide, the DPA’s prompt incorporates comprehensive information about each available tool, including specific usage instructions and the parameters available.

This structured approach helps the DPA refine the plan, select tools and add parameters, taking into account the outputs of prior steps to ensure each subsequent action builds logically on the last.

Additionally, one-shot examples are provided to guide the model in creating accurate specifications, enabling the DPA to deliver a detailed, actionable set of instructions to be executed.

• Function: Expands the high-level plan by specifying the exact tools and parameters needed for each step, ensuring the entire plan is logical and contextually aware of the outputs of all prior steps.
• Input: Receives the high-level plan from HLPA.
• Model Choice: Utilises larger, more capable models to manage the complexity of tool selection and parameter specification. (e.g., GPT4o+, o1, Claude 3.5+, Llama 3+ 70B etc).
• Output: Produces a detailed, contextual JSON plan with specific tools and parameters for each step.

Example Output:

{
  "steps": [
    {
      "step": "Flight Booking",
      "tool": "FlightSearchAPI",
      "parameters": {
        "departure": "London",
        "destination": "Paris",
        "date": "2024-08-01",
        "class": "Economy"
      },
      "output": {
        "confirmation_number": "ABC123"
      }
    },
    {
      "step": "Accommodation Booking",
      "tool": "HotelBookingAPI",
      "parameters": {
        "location": "Paris",
        "check_in": "2024-08-01",
        "check_out": "2024-08-07",
        "guests": 2,
        "reference": "ABC123"// Using output from previous step
      },
      "output": {
        "reservation_id": "HOTEL456"
      }
    },
    {
      "step": "Itinerary Planning",
      "tool": "ItineraryPlanner",
      "parameters": {
        "destination": "Paris",
        "interests": ["museums", "restaurants", "historical sites"],
        "days": 6,
        "hotel_reservation_id": "HOTEL456"  // Using output from previous step
      }
    },
    {
      "step": "Transportation Arrangements",
      "tool": "LocalTransportAPI",
      "parameters": {
        "city": "Paris",
        "options": ["metro", "taxi", "bus"]
      }
    },
    {
      "step": "Packing List",
      "tool": "PackingListGenerator",
      "parameters": {
        "destination": "Paris",
        "duration": 6,
        "season": "summer",
        "activities": ["sightseeing", "dining"]
      }
    }
  ]
}

Rationale: Taking this step out of HLPA makes the planning more specific and enables the planning to be finer and more coherent. This will distribute the cognitive load across different systems, each of them playing to their strengths. By providing comprehensive details and examples in the prompts, one ensures that the DPA generates the specifications with accuracy. Also, during each step, the contextual awareness of previous steps’ outputs can allow the generation of a more coherent and logical plan.

3. Action Agent (AA)

The Action Agent (AA) processes the JSON output from the Detailed Planning Agent (DPA) in a step-by-step sequence. For each step, the AA calls the specified tool using the parameters provided by the DPA and validates/adjusts them based on instructions embedded within the function call. The AA also have temporal memory, so it is able to reference parameters from previous steps, such as using a country ID retrieved earlier to look up flights. This approach helps ensure that each step is executed precisely and efficiently, closely following the detailed plan set by the DPA.

• Function: Sequentially executes each step of the detailed plan.
• Implementation: Utilises simple function calls with predefined tools and parameters. Each function call is provided only one tool, which anecdotally seems to speed things up and reduce variability.
• Safety Feature: Incorporates a small LLM (ChatGPT3.5/4o-mini etc) for parameter validation and handling edge cases.

Example Execution:

{
  "step": "Flight Booking",
  "execution": "FlightSearchAPI.search_flights(departure='London', destination='Paris', date='2024-08-01', class='Economy')"
}

Rationale: This structured approach limits guesswork and can, therefore, bring in a number of errors. Also, embedding a small LLM with function calls for parameter validation and handling edge cases adds robustness without having significant detriments to performance enhancement.

4. Writing Agent (WA)

Finally, the Writing Agent (WA) synthesises the outputs from each completed step into a response. The WA draws on information from each step and follows instructions from its prompt to output the desired narrative in relation to the user's query.

• Function: Synthesises the final response by compiling information from all executed steps.
• Input: Array of responses generated by the Action Agent.
• Implementation: The choice of model depends on the number of responses expected per agent request and the desired presentation style. For simple tasks, use the smallest effective model; for more complex tasks that need to consider lots of data, for example, a larger foundation model will be necessary.
• Output: Depending on your solution needs, the Writing Agent can generate concise summaries, detailed narratives, or structured data outputs, ensuring the final result is coherent and useful.

Example Execution:

Input:
[
  {"step": "Flight Booking", "result": "Flight from London to Paris booked for 2024-08-01." },
  {"step": "Accommodation Booking", "result": "Hotel in Paris reserved from 2024-08-01 to 2024-08-07." },
  {"step": "Itinerary Planning", "result": "Itinerary includes visits to the Louvre, Eiffel Tower, and a Seine River cruise." },
  {"step": "Transportation Arrangements", "result": "Metro and taxi arrangements confirmed for local travel."},
  {"step": "Packing List", "result": "Packing list generated: summer clothes, comfortable shoes, travel documents, and camera."}
]

Output:
Your trip to Paris is all set! Your flight from London to Paris is booked for 2024-08-01. You will be staying in a hotel from 2024-08-01 to 2024-08-07. The itinerary includes visits to the Louvre, the Eiffel Tower, and a Seine River cruise. Local transportation via metro and taxi has been arranged. Don't forget to pack summer clothes, comfortable shoes, travel documents, and a camera. Enjoy your trip!

Rationale: The Writing Agent ensures that the final output is not just a collection of individual responses but a comprehensive narrative that fully addresses the user’s original query based on the information provided.

Why SeqRAG Works in Production Environments

SeqRAG’s design aims to make it well-suited for the rigours of a production environment based on the following key features:

1. Speed: By structuring tasks sequentially and reducing computational overhead, SeqRAG’s approach is designed for production environments where consistency and speed are key.
2. Reliability: The structured, step-by-step approach reduces variability and makes outcomes more predictable — an essential quality for production environments.
3. Debuggability (that’s a word, right?): In case things go wrong, finding and fixing the problems is much easier with SeqRAG’s linear process. Issues can be identified as early as the first step without having to wait for the whole plan to be executed.
4. Understandability: SeqRAG adopts a sequential way of doing things, which, in our experience, has helped make it more understandable for the developers implementing it. While other models involve the interaction of agents, each performing some part of the task, adding considerable complexity and unpredictability to the model, SeqRAG ensures clarity and, hence, easiness in handling. It's easy to test each agent independently for a variety of scenarios.
5. Less Computational Overhead: SeqRAG reduces computation overhead by avoiding continuous replanning.
6. Flexibility: In the modular SeqRAG architecture, these components can easily be swapped or upgraded as better technology becomes available.

Limitations of SeqRAG

Although SeqRAG has a set of advantages, it also has limitations that impact its effectiveness in certain scenarios:

1. Initial Plan Errors: If the plan is wrong at the beginning, there is no method for SeqRAG to automatically correct it. Future research could be spent adding validation steps or heuristic checks, which would catch such errors more quickly than later. However, this weakness could also be seen as a strength as it makes it easy to identify whether or not an execution of SeqRAG is likely to be successful.
2. Limited Adaptability: It is less able to handle unexpected information or changes than more adaptive approaches.
3. Dependency on Initial Planning Quality: It is, to a certain degree, dependent upon the quality of the initial HLPA plan.
4. Risk of Overspecialization: It may be less versatile for general-purpose problem-solving than more flexible approaches.

Areas for Further Research and Optimisation

1. Dynamic Model Selection: An adaptive model selection tool based on query complexity chooses bigger models for complex plans and smaller ones for simpler operations, finding a balance between relevance and speed.
2. Parallelisation: For example, creating more than one top-level plan in parallel and then selecting the best out of them. Also, the steps in the detailed plan, which are independent of each other, could be performed parallel to save time.
3. Feedback Loops: Studying the possibility of using execution feedback in the planning phases, but this is too slow now and introduces too much non-determinism, hence making debugging burdensome.
4. Caching and Transfer Learning: The effective storing and reusing of common plan elements when similar queries are executed could bring many performance benefits. As could improvements in prompt caching.
5. Plan Validation: The introduction of a validation step for high-level and detailed plans may increase its reliability. It could be performed separately, either by an agent or through heuristic regex checks, to ensure that the plan is feasible to execute and complete.

The Evolution of AI Planning and SeqRAG’s Place Within It

The broader advancements in AI-assisted planning have informed SeqRAG’s development and are worth being aware of.

Frameworks like PlanRAG, LLM-DP, LLM+P, ReAct/Reflexion, and Microsoft’s AutoGen each represent distinct approaches to complex task management, serving as both inspiration and points of differentiation for SeqRAG.

PlanRAG emphasizes adaptive retrieval-based planning, adjusting dynamically through each stage, whereas SeqRAG opts for upfront planning to prioritize speed and consistency. Similarly, LLM-DP combines LLMs with symbolic planners for tasks that involve physical interactions, diverging from SeqRAG’s focus on structured information retrieval. The LLM+P and ReAct/Reflexion approaches add layers of adaptability and reflection, handling unpredictable environments, while SeqRAG’s design focuses on minimizing variability.

Finally, AutoGen really exemplifies flexible, multi-agent workflows adaptable to diverse applications and served as the initial inspiration for SeqRAG’s specialised agent roles.

SeqRAG’s purpose is not to replace these varied frameworks but to serve as an efficient, reliable option for scenarios that benefit from a structured, step-by-step approach — making it hopefully suitable for predictable, multi-step tasks in production environments.

SeqRAG provides one approach to handling complex, plan-based queries, focusing on efficiency, consistency, and specialisation through sequential steps. We hope this framework offers a starting point for further experimentation in developing reliable, production-ready AI agents.

Dive into our GitHub repo, get your hands dirty with the code, and join the Agent Revolution.

Happy agent building, everyone!

Special thanks to Henry Cleland, Erik Schwartz, Colin Ke Han Zhang. We worked together for many years on the ideas that eventually became SeqRAG, and they deserve equal credit if anything above is useful. Thank you guys!

Disclosure of Affiliation

As one of the creators of SeqRAG, developed in collaboration with my team at Elsevier, I am excited to share our approach to making complex AI workflows more accessible. This project is our way of contributing to the broader AI community, which has inspired our work. It offers a solution we hope others will find both practical and educationally useful.

SeqRAG: Agents for the Rest of Us was originally published in TDS Archive on Medium, where people are continuing the conversation by highlighting and responding to this story.

Warning for my regular readers: The following is experimental writing compared to my usual product management and technical musings. Proceed with curiosity.

Eating the first cream cracker from a packet is one of lives greatest pleasures.
No matter how bad, stressful, joyful or purposeful life gets, biting into that first pristine cracker from the packet somehow always recalibrates my senses.
There is palpable joy at the tipping point before it reaches my lips; it forces a reset of my senses.

Finding joy in a cream cracker is the perfect metaphor for the human condition: endlessly chasing crumbs of happiness in a world that offers us the whole cake.
I don’t know why such a simple thing gives me such momentary pleasure.
There is no hunger, nostalgia or Proustian moment where I can ground these feelings.
But we all experience them—small, seemingly mundane moments that drive so much unappreciated pleasure—moments of quiet celebration.
A moment to reflect, even briefly, on the beauty of small things.

Your cream cracker may be chocolate, a walk, a hand wave, the sound of a gong, or an aeroplane flying overhead.
They are things that scream an inverse correlation between an activity’s usefulness and its enjoyability; they should not command such power.
Yet, these small triggers can take us out of ourselves, make us fully present in a moment, and life worth living.

In its simplicity, the cream cracker is a symbol of rebellion against modern life’s dictum that more is better.
Yet, in this act of defiance, we are still trapped within the same system that dictates our desires.

Some may say that by elevating the consumption of a cream cracker to an event of joy, we reveal our own complicity in the ideological machinery that keeps us perpetually unsatisfied.
A fetishisation of ‘authentic’ experiences in an inauthentic world.
It’s a trick that we play on our own minds that we are so evolved in our thinking that we can rise above the noise of distractive digital modernity and appreciate such moments.
That in our celebration of the first cream cracker, we see the tragedy of modernity: the profound moments of life are not lived but eaten.
Still savant to the capitalist consumption machine that produced it.

That all said.
Enjoy that damn cream cracker.

“Work while you can.”

That was my mantra going into 2023 — the year I first truly contemplated the notion of being replaced by a machine. Since then, things took a rather unexpected turn. Now, rather than being a looming threat, AI has since emerged as a kind of companion.

This pivotal shift occurred around mid-year amidst a particularly challenging product decision and escalating team conflict. Together with my team, we turned to AI, seeking an innovative solution. We fed ChatGPT the parameters and arguments of our dilemma, initiating a multi-perspective role play with the AI. This exercise, unfolding over an insightful hour, unveiled a spectrum of potential solutions.

Surprisingly, implementing these AI-generated strategies led to a significant transformation. Relationships within the team were not just mended; they were strengthened, and the trajectory of our project shifted dramatically for the better. AI, I realised, is more than a mere writing tool. It serves as a mediator, a reflective mirror, casting light on our ideas and biases that guide us to better outcomes.

My experimentation continued. I learned to engage in Socratic dialogues with AI. I presented information and scenarios and got it to critique my approaches, forcing me to think differently about my assumptions. My professional work improved — I even created a toolkit for product managers to share my approaches. My personal development improved, and it even helped me overcome some anxiety and develop new mental frameworks.

So, as we step further into this evolving relationship with AI, let’s do so with caution, purpose, stewardship and a touch of humour. Here’s to 2024, where we don’t merely coexist with AI but thrive alongside it.

I hope you find something inspiring in my reflections today.

AI makes us more creatively competitive

• The value of LLMs is not in their text generation, but rather, in their ability to make any idea more convincing.
• Machines are particularly well suited to providing input for creativity, as their connections are more diverse and thus have a less compartmentalised view of the world than us humans.
• Getting an LLM like ChatGPT to ask me questions about any given idea, statement, report or email with the goal of those responses to improve my work is a game changer for productivity. Think of it like a socratic workout for the mind — but applied to everything!

“My students and I almost always got more effective results when we wrote original drafts on our own and only used AI as a tool to help us develop ideas, rather than having AI generate a passage that we then attempted to spruce up.” — Todd Mitchell

• While it’s natural to feel concerned about AI’s potential implications on various professions, it’s essential to remember that our adaptability and ingenuity in knowledge work have always led to new opportunities for humanity.

Reading List
The Art of Obligation in the Age of AI
Breaking academic barriers: language models and the future of search
AI won’t steal your job, people leveraging AI will
Why AI Will Save the World
AI and the Age of the Individual

What Writers Need to Know to Be Better Than AI

For the best outcomes, work to make the other person’s argument better

• Conflict is the catalyst of creativity.
• Actively improving another’s argument against one's own argument is an exercise in intellectual humility and honesty. It’s a powerful tool for self-reflection and building rapport, even if you don’t end up changing your mind.
• If you succeed in making your opponent’s opinion, should it prove true, or seem distinctly to their own interest, your opponent will drop it like a hot potato.
• Constantly communicate the ‘why’ behind your decisions, whether you decide to act on the feedback or stay the course.
• Aim to build a process that strips away ego and targets the best outcome. As Jim Collins once noted, “Good outcome with bad process reinforces bad process,” so lets aim for the best process possible.
• The chief enemy of good decisions will always be a lack of sufficient perspectives on a problem.

“Conflict is the beginning of consciousness” — M. Esther Harding pointed out.

• Studies show the best work cultures try to get issues out in the open. Active debate is what leads to the best brainstorming and top notch ideas. Best example? Wikipedia. — Eric Barker
• A general principle underpinning all our processes is that important decision should not be made alone. Good team leaders know when to consult others and seek perspectives different from theirs.

Reading List
A product manager’s 56 best (and stolen) insights from 2021
The Work Required to Have an Opinion
How to have better arguments online
New Neuroscience Reveals 7 Secrets That Will Make You Emotionally Intelligent
Schopenhauer’s 38 Ways to Win an Argument Are

Being opinionated is the force that drives change

• Rule: Be opinionated, embrace criticism, and improve the arguments of others.
• The cardinal rule for leading a Tiger Team as a product manager is to be unapologetically opinionated about what a “good” outcome looks like.
• Strong leaders engage their critics and make themselves stronger. Weak leaders silence their critics and make themselves weaker. — Le Bon

“The only thing worse than being blind is having sight but no vision.” — Helen Keller

• As a product leader, it’s your job to make an opinionated decision and decide whether to optimize for learning or for first-time quality.
• Adopting an opinionated stance provides crucial clarity, streamlining complex decision-making and ensuring alignment across the team.
• Being opinionated isn’t just about setting a course; it’s about wholeheartedly owning the mission.
• Strong leadership doesn’t mean steamrolling others. Hold strong opinions, but be willing to change when presented with compelling counterarguments — Le Bon

“Instead of spending our lives imagining what things might be like, we’re often better off trying them out and seeing.” — kyleschen, What Do You Want to Want?

• If you don’t notice your strongly held opinion, you can’t challenge it — Michael Ashcroft
• You can cut all the chains holding you down, but if you have nowhere to go, then what was the point? To quote Kierkegaard, “anxiety is the dizziness of freedom.”
• Because you’re making opinionated decisions, you have to be accountable for your opinions.
• Remember though — you are your values, not your beliefs. When your sense of self comes from the opinions you hold, it makes it far more difficult to change them.
• And finally, because you’re making opinionated decisions, you also have to be accountable for those strong opinions.

Reading List

Taming the tiger: a product team’s guide to risky projects
So you want to build a search engine
Brainwashed by design
You Can Only Respond to What You Notice
Think Again (Adam Grant)

Planning is an Antidote to Despair

• The mere act of planning events to look forward to acts as a lifeline. Whether it’s a trip or a simple dinner, it creates an anchor, making you question, ‘What’s the rush to give up when there’s something good on the horizon?’
• “We drum our fingers, looking ahead to the next event, longing for some kind of future amusement and wishing we could leap over the days in between.” — Jack Maden on Seneca

“A bias for action is the best way to ensure that we stay connected to reality; which is, in turn, our best defense against wants we don’t want to want.” — kyleschen, What Do You Want to Want?

• Planning is also paradoxically risky for people in precarious situations since it requires you to commit to a future course of action, one that could come with financial penalties if you have to change your mind; but also, penalties if you don’t plan, in the form of fewer options and higher prices. — Adrian Hon
• Inaction creeps in when we let down our guard. We need to condition ourselves to action if we want a life worth living. — Joseph Wells

Reading List
Without Reservation
Seneca on Coping With the Shortness of Life
Algorithms to Live By: The Computer Science of Human Decisions (Brain Christian)
Action Is Everything

Innovation is when you can’t name a thing

• If you want to invent things guide yourself to places where you run out of names for things.
• Give it a name and believe in it, and try to make it come true. That’s progress. — Paul Ford
• The best way to fail at inventing something is by making it somebody’s part time job.

“Do not waste the little span allotted to you producing only work intended for the moment rather than for posterity.” — David Hart

• Innovation thrives on reinventing and reusing the old.

“Progress means a 92 year old who did not die today, a boy who was not robbed on his way to school, a 12 year girl who is not married to a 30 year old man.” — Kevin Kelly

• I’ve seen the most impactful innovations occur not in isolation, but at the crossroads of invention and application. This year has reinforced my belief that those who engage directly with the real world often find the most success.

Reading List
Innovation is slowing down
Why Humans Are So Bad at Seeing the Future
Kevin Kelly The Case for Optimism
Working Backwards (Colin Bryar)

Showing gratitude and creating obligation improves mental health

“Daily acts of gratitude, both spoken and silent release … mental pressure and subtly shift your approach to interactions, often with unexpectedly positive outcomes.” — Julian Baggini

• Say thank you, open doors for others, don’t interrupt and so on. These little acts help to lubricate our social interactions, making everyone’s lives more pleasant.
• When gratitude is expressed publicly, it can also enhance one’s reputation in the workplace and in one's community and life.
• It’s important to encourage and be grateful, even when you can’t immediately act on it.
• Create a template with crucial questions, such as “What’s top of mind right now?”, “What would make today great?”, “What am I thankful for?” and spending the first 10 minutes of each day answering them.

“Moving toward values is about reconnecting our day to day activities to the things that matter most to us.” — Casey Rosengren

• Show me your calendar, and I’ll show you what you value.
• Presence is better protected by obligation than intention.
• Embracing obligations can allow us to connect with our uniquely human traits and forge meaningful connections in a world filled with more AI and automation.
• Psychologist Michael Tomasello’s research on the moral psychology of obligation highlights how historically, obligation distinctly motivates people and is a force with insistent nature that bonds us with others and encourages us to honour agreements.
• The fact that someone took the day off and decided to spend it with you doing something fun suddenly becomes meaningful. — Simone de Beauvoir
• Cultivating obligations need only start with a calendar and a notepad.

Reading List
The Art of Obligation in the Age of AI
Day in the life of an obsessive note taker
How To Do Hard Things
Manners matter — philosophy tells us why
How to Criticize Coworkers
Against Ikigai

Metrics are the biggest curse on product development

• According to Gartner research, 96% of customers who identify a product or service experience as high effort become disloyal, compared to only 9% who have a low effort experience.
• For any metric of success, we have to be cautious as that metric may gradually become less useful over time, especially as we get better at manipulating it.
• It’s crucial to recognise that not all usage metrics are positive. You see, sometimes users don’t have a choice — they’re stuck using a product that drives them crazy, whether it’s because their bank has a great interest rate, their family only uses Facebook Messenger, or their company makes them use Workday. These are what we like to call “Hidden Detractors”, and that is why it's so important to also measure how satisfied users are in how you are helping to solve their problem.
• Embracing quantifiable metrics prevents us from getting lost in the intellectual maze of success metrics. It’s a straightforward process — define the problem, understand your user base, quantify engagement, and measure satisfaction. This method proves universally applicable, slicing through potentially endless debates about OKRs.
• Traditional business-level metrics, such as NPS (Net Promoter Score), often fall short in providing a clear picture of a product’s performance. They might capture sentiment but miss the nuances of user experience and engagement.

Reading List
‘Working Backwards’ to finally make remote work… work
So you want to build a search engine
Why Companies Stop Innovating
Why Humans Are So Bad at Seeing the Future
How to Write a PRD That Actually Helps You Build Products

Failure is the most important part of life

Embrace failure as an integral part of life’s journey. Treating failures as lessons rather than burdens has offers insights into your character and resilience.

• But if you know you are on a path of improvement that will definitely lead somewhere interesting, setbacks aren’t as threatening, and success along the way isn’t something to desperately cling to. — Nathan Baschez
• In the event that you, as a leader, share a dissenting opinion that proves to be incorrect, it’s enormously valuable to moral and team growth to call out your mistake. — Natalie Rothfels

“When you don’t have any chains and you accept the aimlessness of things the flip side of that is that now you can go anywhere.” — Emil Cioran

• Researchers have found that embracing a life full of challenge, though it involves pain and effort in the short term, ultimately leads to greater happiness and flourishing in the long term.

Reading List
‘Working Backwards’ to finally make remote work… work
FRAME Facilitate Better Product Reviews in 5 Steps
Book Review Mastery
What Do You Want to Want?
Episode 156 Emil Cioran Part 2 Failure and Suicide
How to Find Clarity When You’re at a Career Crossroads

To make AI more aligned, you first need to understand that it will never see the world as you do

• Our creations are, in many ways, a reflection of us and, therefore, likely to be as flawed.
• Because AI’s being trained on average human work, they exhibit the average human’s biases, prejudices, weaknesses, and vices, which may become permanently encoded.
• Just as no single individual possesses all the answers, no AI, no matter how advanced, will ever encapsulate every perspective — much like a calculator can’t simultaneously encompass every possible calculation.

“The truth is, any system that must be perfect to be reliable is deeply flawed.” — Tiago Forte

• Even if AI develops a form of consciousness, it’s uncertain whether it will value or perceive the world as we do. Our connection to each other, our shared obligations — that’s uniquely human.
• Rather than predicting every move in the chess game of life, our AI should focus on understanding the current state of the board and justifying its actions.
• A critical aspect often overlooked in ‘2001: A Space Odyssey’ and should act as a warning for future smart AIs, is that the supercomputer, the HAL 9000 wasn’t inherently evil or malfunctioning. Designed for truthfulness, precision, and obedience, HAL encountered a moral dilemma: reconciling contradictory directives. The machine grappled with keeping the mission’s true nature a secret from the astronauts while maintaining honesty in its interactions, and that is what drove it to insanity. “Gulp”

Reading List
How to Fix a HAL 9000
If We Succeed
What Writers Need to Know to Be Better Than AI
Episode #184 — Is Artificial Intelligence really an existential threat?
Interview Kevin Kelly, Editor, Author, and Futurist

Happy 2024 EVERYONE!

A product manager’s 69 insights from 2023 was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

Imagine being handed a ‘mission impossible’: You have just five weeks to build a groundbreaking generative AI product, and you’re doing it with a team that’s never worked together before. The odds are stacked against you — according to Gartner, a mere 53% of machine learning projects progress from prototype to production. Even organisations seasoned in AI would find this daunting.

With over a decade of product management experience under my belt, I thought I was ready for anything. But this mission was different. It turned my long-standing product principles upside down and fundamentally changed the way I’ll approach product management from now on.

Through this challenging experience, I learned three unorthodox but invaluable lessons that I like to call my "Tiger Team Tactics". Although these ideas are not new on their own, when used together, I’ve found them to be a productivity game-changer. Whether you are an experienced product manager or taking on a leadership role for the first time, these tactics will empower you to lead your high-stakes projects to success.

Join me as we dive into these hard-earned insights and tools for corralling your own ‘team of tigers.’

Tiger Teams and the AI Goldrush

Generative AI is sparking a modern gold rush reminiscent of the dot-com boom of the 1990s. Even McKinsey estimates that generative AI could soon contribute up to $4.4 trillion annually to the global economy. Naturally, such estimates have piqued the interest of many organisations.

In this high-speed landscape, the traditional playbook for incremental product development is being tossed aside in favour of rapid, game-changing outcomes that directly meet a clear user need. Enter Tiger Teams, a high-stakes, high-reward strategy birthed during the Apollo 13 era. These specialised squads are becoming the go-to approach for companies eager to seize immediate opportunities in this surging AI revolution.

A Tiger Team is a melting pot of diverse skill sets — from engineers and data scientists to analysts, marketing and sales professionals — united under a single mission. Traditional frameworks of product management often fall short in this setting because they focus on lengthy problem identification, market validation and iteration. However, in order to gain a competitive edge, it is important to prioritise harnessing specialised technologies or ideas to resolve existing problems. This requires a shift from managing products to managing people and strategy.

Leading a Tiger Team demands a different playbook, one that favours swift execution and competitive positioning over traditional, extended pre-planning and poses a unique combination of challenges:

1. Ambiguity: Pioneering an untested concept with unproven market demand.
2. Velocity: Achieving ambitious objectives on a tight deadline without compromising quality.
3. High Risk: Confronting the ever-present risk of failure.
4. Team Dynamics: Navigating complex interpersonal and professional relationships under stress.
5. Stakeholder Equilibrium: Striking a delicate balance between team expectations and upper-management imperatives.

It's not just a technical challenge; it’s an emotional odyssey. You shoulder not just the aspirations of your team but also the high stakes set by your organisation. So, how do you build trust, maintain productivity, and encourage open dialogue amid such pressure with the hope of success?

In the following sections, I’ll unpack the three Tiger Team Tactics — being opinionated, critical feedback and steel manning arguments — that helped me successfully steer a Tiger Team and launch our AI product within an ambitious six-week timeframe.

Tiger Tactic 1: Embrace Opinionated Leadership

Your First Question: What Does “Good” Look Like?

The cardinal rule for leading a Tiger Team as a product manager is to be unapologetically opinionated about what a “good” outcome looks like.

It's about having a clear vision: a way to help you and the team decipher whether a decision is good or bad, based on whether it gets you closer or further away from where you want to go. It doesn’t mean being inflexible or resistant to change; instead, what “good” looks like sets the stage for crucial conversations and decisions that must be made in a time-sensitive environment.

See, in fast-paced, high-stakes settings like a Tiger Team, lengthy brainstorming sessions and extended workshops are luxuries you can’t afford. Adopting an opinionated stance provides crucial clarity, streamlining complex decision-making and ensuring alignment across the team.

And being opinionated isn’t just about setting a course; it’s about wholeheartedly owning the mission. This is important because the heightened sense of responsibility eliminates ambiguity and cultivates a fertile environment for actionable feedback.

“The only thing worse than being blind is having sight but no vision.” — Helen Keller

Crafting Your Opinionated Vision: A How-To Guide

With traditional product development, the focus starts with identifying a valuable problem, followed by seeking a solution.

However, as mentioned, an opinionated product leader flips this paradigm: Start by defining what a successful outcome looks like (‘What good looks like’), and let that vision guide your problem-solving efforts. Below are practical steps I’ve found effective for crafting such an opinionated vision:

1. The Press Release: They say you can’t think clearly until you can write clearer. Start with a hypothetical press release you’d like to publish upon project completion. Borrowed from Amazon’s ‘working backwards’ approach, this technique crystallises your vision in clear, concise terms. Share it with a small group, then refine until that vision resonates with and impresses your audience. Fun tip — Slides have their use, but nine times out of 10, the answer to getting buy-in from busy execs is a one-page memo.
2. The Product Six-Pager: Expand on your vision with a comprehensive product document outlining the problem, inspiration, proposal, guiding principles, success metrics, and FAQs. This is more than a checklist; it serves as your strategic blueprint and should be a basis for broader constructive criticism. Strong six-pagers don’t just make their case; they anticipate counterarguments, points of contention, or statements that might be easily misinterpreted.
3. Visualising the Vision: Many recommend low-fidelity visuals initially, but I argue for starting with high-fidelity, especially if you have a skilled designer on your team. High-quality mockups lead to more honest and effective conversations from the get-go.

Don’t let your vision document become an overlooked email attachment. Schedule a focused team discussion to discuss it: allocate 15 minutes for silent reading, followed by a structured round-robin feedback session. Use this feedback to fine-tune your vision, transforming it into a “shared vision” that galvanises the entire team.

Takeaways

Strong leadership doesn’t mean steamrolling others. Hold strong opinions, but be willing to change when presented with compelling counterarguments:

• Kick off your project by sharing a clear, opinionated vision and then explicitly ask for feedback.
• Utilise both a press release and a six-pager as foundational guiding documents.
• Focus on “good” outcomes rather than just problems to solve — trust me, by doing so, you’ll naturally identify the right problems to solve.

Risks and Caveats:

• Team Alienation: Being too opinionated can alienate team members, undermining collaboration and morale. This is a pitfall I’ve experienced firsthand. Keep an eye out for this (Rules 2 and 3 should help here).
• Confirmation Bias: A strongly held vision might cause you to seek out only the information that supports your views while dismissing valid concerns or data that contradicts them.

Regardless, I would recommend you seize every opportunity to pose the question — to yourself and your team — “What does ‘good’ look like?” and formulate a strong perspective based on that inquiry.

Coming up, we’ll delve into the critical skill of actively seeking and incorporating feedback to ensure a balanced and practical approach.

Tiger Tactic 2: Actively Seek Criticism

Jim Collins once noted, “Good outcome with bad process reinforces bad process.” To address this in product management, feedback is not just welcomed — it’s vital. While strong opinions can streamline decisions (as outlined in Rule 1), it’s crucial to counterbalance this by actively seeking criticism.

Being overly opinionated risks creating a divisive, ‘you’re either with me or against me’ environment, which can stifle creativity and overlook valuable perspectives. To break free from this echo chamber, we need to establish structured mechanisms for gathering constructive feedback.

Daily “Superfriends” Meeting: Coffee and Candor

The most impactful tool I have for candid feedback is the “Superfriends” meeting. This 15-minute daily stand-up involves key team members like tech leads, designers, and analysts. The sole question asked to each member is, “What is top of your mind right now?” Everyone gets a minute or two to share their thoughts.

The objective of these meetings is not to solve all problems but to establish a continuous feedback loop. These quick check-ins not only gauge the team’s pulse but also preempt potential issues, thereby enhancing decision-making.

Sprint Retrospectives: Your Regular Reality Check

While many product teams are familiar with the fortnightly Agile Retrospectives, their importance, especially in time-sensitive settings, can’t be overstated. DON’T BE TEMPTED TO SKIP THEM! These retrospectives still play a crucial role in team cohesion. They offer a platform for active listening and discussion. You don’t just collect feedback; you honour it by integrating insights into future sprints.

Team Surveys: Empowering the Introverted

Not everyone is comfortable voicing their opinions in a group of peers or superiors. Anonymous surveys offer a channel for even the most introverted to share their thoughts without inhibition. These unfiltered inputs can reveal unexpected but vital insights.

Turn Criticism into Goals

Feedback can often feel like criticism, and nobody loves a critic. Yet, critics can be the sculptors of a masterpiece. Serious concerns that emerge from feedback should be candidly discussed and consciously integrated into team goals and objectives. While you may not act on every piece of feedback, it’s essential to at least acknowledge and consider it.

Transparency in Decision-Taking

Constantly communicate the ‘why’ behind your decisions, whether you decide to act on the feedback or stay the course. This cultivates a culture where criticism is not merely tolerated but actively sought and helps more team members feel connected to the project and its outcomes.

Takeaways

By systematising feedback mechanisms, we aim to build an environment where criticism catalyses ongoing refinement. Feedback is important for identifying and addressing bad habits and ensuring that it is delivered in the right context to avoid feeling like unwarranted criticism.

Risks and Caveats:

• Analysis Paralysis: Too much feedback can overwhelm decision-making processes, slowing progress and causing indecision.
• Over-Reliance on Consensus: While feedback is valuable, a product manager must sometimes make unpopular decisions. Striving for universal agreement can weaken your strategy’s effectiveness — remember to stay opinionated and have solid reasoning to back it up.

In the next section, I will explore how to master conflict resolution through the Steel Man Technique, ensuring that criticisms serve as stepping stones rather than stumbling blocks.

Tiger Tactic 3: Master Conflict Resolution with the Steel Man Technique

“Conflict is the beginning of consciousness,” as psychoanalyst M. Esther Harding pointed out. Leading a Tiger Team involves more than just decision-making; it requires mediating high-stakes situations without causing division.

The secret? Emotional intelligence coupled with an approach for managing diverse opinions.

Feedback is a Double-Edged

A Tiger Team is both passionate and opinionated — qualities that can be both an asset and a potential liability.

While a rich diversity of viewpoints can ignite innovation, it can also become a disruptive force if not carefully managed. It is critical to create a safe space where team members feel their opinions are valued.

The Steel Man Technique: Intellectual Honesty

The key is to approach the argument with the intention of changing your own mind if necessary.

For this reason, I favour the Steel Man Technique for managing feedback. The approach is simple — make the other person’s argument better than they can. Once improved, present both your own argument and the bolstered version of your opponent’s stance back to the relevant group members. If you can improve upon your opponent’s argument in a way that defeats your own, it’s time to change your mind.

This may initially seem counterintuitive: why argue for a viewpoint you fundamentally disagree with? The answer is simple yet profound. By articulating your opponent’s argument in the most robust way possible — steel-manning it — you accomplish several things simultaneously:

1. Intellectual Integrity: You demonstrate that you understand the critique or opposing viewpoint deeply enough to improve upon it. This is intellectual honesty in action, an indispensable virtue in any team environment, especially a high-stakes one.
2. Building Trust: When you can articulate a team member’s viewpoint better than they can, you earn their trust, a fundamental element for team cohesion. Challenge yourself to do this consistently.
3. Critical Self-Assessment: If the ‘steel-manned’ argument stands up better than your own, it’s a red flag that you need to reconsider your stance.
4. Clear Communication: The technique pinpoints exactly where agreements and disagreements lie, making it easier to find a mutually agreeable path forward.

Organisational psychologist Adam Grant highlights that when managed effectively, conflict can serve as a catalyst for creativity. The goal is to help both parties become their most authentic selves and achieve satisfaction rather than compromise.

In my own experience with our own AI product, implementing the Steel Man Technique has been pivotal. It not only helped me navigate challenging conversations but also fostered a better team culture.

Case Study from our AI Product

To illustrate the Steel Man Technique in action, let me share a recent experience from our AI project. With only two weeks until launch, I found myself at odds with our Head of UX. The issue at hand? A critical change in user flow that would also be challenging from an engineering standpoint. I was convinced my approach was correct; my colleague strongly disagreed.

Rather than dismissing their point of view, I took the time to understand and amplify it. I pushed myself to articulate their argument more convincingly than they had and, in doing so, gained new insights. It was bitter medicine to take (and one late night), but I needed it. The exercise prompted a beneficial change in our course of action, executed just in time for the project’s launch. The team understood the rationale for this last-minute shift and trusted the decision-making process.

The Steel Man Technique has helped me tackle confrontations both small and large. It facilitates course corrections without escalating conflicts, empowers team members to influence project decisions, and reinforces a culture that highly values opinions. Most importantly, this transparent method has built trust in decision-making, as everyone knows their input will be genuinely considered.

Takeaways

Steel Manning is more than just a tool for debate; it’s a framework for fostering intellectual honesty, building trust, and encouraging open communication within your team.

By embracing this technique, you not only challenge your own viewpoints but also create an environment where everyone’s perspectives are valued and scrutinised constructively.

Risks and Caveats:

• Decision Complexity: While steel-manning is helpful, it can lead you to overanalyse, adding complexity and slowing down decision-making. Set time limits for decisions, acknowledge that you may not have complete information, and commit with the team, understanding you can adjust later if you are wrong.
• Risk of Appeasement: If you find yourself continually steel-manning and adopting others’ viewpoints, you may deviate too far from your original strategy, goals, or vision. When this happens, work to remember what “good” looks like and be opinionated about it.

Overall, the key is to approach the argument with the intention of changing your own mind if necessary and to focus on helping both parties reach a mutually satisfactory outcome.

Leading a Tiger Team is exhilarating but demanding. The landscape is rife with unique challenges, but navigating them successfully is immensely rewarding.

Equip yourself with three indispensable Tiger Team Tactics — be opinionated yet open, proactively seek criticism, and embrace intellectual honesty during conflicts — to achieve the best outcomes for your team.

Here’s a quick recap of the rules to help you succeed:

1. Be Opinionated About What “Good” Looks Like: Set a clear direction and mission for the team.
2. Actively Seek Criticism: Establish a culture and system where feedback is encouraged and integrated into your strategy.
3. Apply the Steel Man Technique: Turn conflicts into constructive opportunities by making the arguments of others better than your own.

These aren’t just theoretical insights; they’re my hard-earned lessons from my experiences in developing AI products. Employing these principles does more than enhance your leadership skills; it uplifts your entire team, propelling your project toward defining what ‘good’ truly means.

Remember, leading a Tiger Team isn’t about winning battles of will; it’s about steering your team to a collective win for your users. In the end, the minor disputes and intense discussions will fade from memory. All anyone will remember is what you delivered.

Happy Tiger Taming, everyone!

Bibliography

Articles and books that influenced the development of my Tiger Team Tactics:

• Why Team Building & Development Is an Important Competitive Edge, Bryan Kitch
• Why Companies Stop Innovating, Steve Andriole
• Aggregated knowledge from a small number of debates outperforms the wisdom of large crowds, Joaquin Navajas, Tamara Niella, Gerry Garbulsky, Bahador Bahrami & Mariano Sigman
• Think Again, Adam Grant
• Working Backwards: Insights, Stories, and Secrets from Inside Amazon, Colin Bryar, Bill Carr
• The Neuroscience of Goal-Setting and Its Impact on Your Culture, Lewis Kallow

Taming the tiger: a product team’s guide to risky projects was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

The Next Frontier of Search: Retrieval Augmented Generation meets Reciprocal Rank Fusion and Generated Queries

Having explored search technologies for almost a decade, I can honestly say nothing has been as disruptive as the recent rise of Retrieval Augmented Generation (RAG). This system is revolutionising search and information retrieval using vector search with generative AI to produce direct answers based on trusted data.

In my search projects, experimenting with RAG has led me to consider its potential enhancements; I believe RAG is still too limited to meet users’ needs and needs an upgrade.

Don’t get me wrong, RAG is excellent and is absolutely a step in the right direction for information retrieval technologies. I’ve used RAG since the advent of GPT-2 in 2021, which has significantly helped boost my productivity when looking for valuable information from my own notes or work documents. RAG has many advantages:

• Vector Search Fusion: RAG introduces a novel paradigm by integrating vector search capabilities with generative models. This fusion enables the generation of richer, more context-aware outputs from large language models (LLMs).
• Reduced Hallucination: RAG significantly diminishes the LLM’s propensity for hallucination, making the generated text more grounded in data.
• Personal and Professional Utility: From personal applications like sifting through notes to more professional integrations, RAG showcases versatility in enhancing productivity and content quality while being based on a trustworthy data source.

However, I’m finding more and more limitations of RAG:

• Constraints with Current Search Technologies: RAG is limited by the same things limiting our retrieval-based lexical and vector search technologies.
• Human Search Inefficiencies: Humans are not great at writing what they want into search systems, such as typos, vague queries, or limited vocabulary, which often lead to missing the vast reservoir of information that lies beyond the obvious top search results. While RAG assists, it hasn’t entirely solved this problem.
• Over-Simplification of Search: Our prevalent search paradigm linearly maps queries to answers, lacking the depth to understand the multi-dimensional nature of human queries. This linear model often fails to capture the nuances and contexts of more complex user inquiries, resulting in less relevant results.

So, what can we do to address these issues? We need a system that doesn’t just retrieve what we ask but grasps the nuance behind our queries without needing ever-more advanced LLMs. Recognising these challenges and inspired by the possibilities, I developed a more refined solution: RAG-Fusion.

Why RAG-Fusion?

• Addressing Gaps: It tackles the constraints inherent in RAG by generating multiple user queries and reranking the results.
• Enhanced Search: Utilises Reciprocal Rank Fusion and custom vector score weighting for comprehensive, accurate results.

RAG-Fusion aspires to bridge the gap between what users explicitly ask and what they intend to ask, inching closer to uncovering the transformative knowledge that typically remains hidden.

Embarking on this journey with RAG years ago, I regret not sharing those initial experiments. But it’s time to make amends. Let’s delve deep into the technical intricacies of RAG-Fusion.

A Deep Dive into the Mechanics of RAG-Fusion

Tools and Tech Stack

FYI, for those of you who just want to see code and get straight into playing with RAG-Fusion, check out the GitHub repo here.

The foundational triad of RAG Fusion is similar to RAG and lies in the same three key technologies:

• A general-purpose programming language, often Python.
• A dedicated vector search database, such as Elasticsearch or Pinecone, steering the document retrieval.
• A potent large language model, like ChatGPT, crafting the text.

However, unlike RAG, RAG-Fusion differentiates itself with a few additional steps — query generation and a reranking of the results.

RAG-Fusion’s Workflow:

1. Query Duplication with a Twist: Translate a user’s query into similar, yet distinct queries via an LLM.
2. Vector Search Unleashed: Perform vector searches for the original and its newly generated query siblings.
3. Intelligent Reranking: Aggregate and refine all the results using reciprocal rank fusion.
4. Eloquent Finale: Pair the cherry-picked results with the new queries, guiding the large language model to a crafted output that considers all the queries and the reranked list of results.

Let’s walk through each of these steps in more detail.

Multi-Query Generation

Why Multiple Queries?

In traditional search systems, users often input a single query to find information. While this approach is straightforward, it has limitations. A single query may not capture the full scope of what the user is interested in, or it may be too narrow to yield comprehensive results. This is where generating multiple queries from different perspectives comes into play.

Technical Implementation (Prompt Engineering)

The use of prompt engineering is crucial to generate multiple queries that are not only similar to the original query but also offer different angles or perspectives.

Here’s how it works:

1. Function Call to Language Model: The function calls a language model (in this case, chatGPT). This method expects a specific instruction set, often described as a “system message”, to guide the model. For example, the system message here instructs the model to act as an “AI assistant.”
2. Natural Language Queries: The model then generates multiple queries based on the original query.
3. Diversity and Coverage: These queries aren’t just random variations. They are carefully generated to offer different perspectives on the original question. For instance, if the original query was about the “impact of climate change,” the generated queries might include angles like “economic consequences of climate change,” “climate change and public health,” etc.

This approach ensures that the search process considers a broader range of information, thereby increasing the quality and depth of the generated summary.

Reciprocal Rank Fusion (RRF)

Why RRF?

Reciprocal Rank Fusion (RRF) is a technique for combining the ranks of multiple search result lists to produce a single, unified ranking. Developed in collaboration with the University of Waterloo (CAN) and Google, RRF, in the words of its authors, “yields better results than any individual system, and better results than standard” reranking methods.

By combining ranks from different queries, we increase the chances that the most relevant documents will appear at the top of the final list. RRF is particularly effective because it doesn’t rely on the absolute scores assigned by the search engine but rather on the relative ranks, making it well-suited for combining results from queries that might have different scales or distributions of scores.

Typically, RRF has been used to blend lexical and vector results. And although that method can help make up for the lack of specificity of vector search when looking up specific terms like acronyms, for example, I’ve been unimpressed by the results, which tend to be more of a patchwork of multiple result sets as the same results rarely come up for the same query for lexical and vector search.

Think of RRF as that person who insists on getting everyone’s opinion before making a decision. Only in this case, it’s not annoying but helpful. The more, the merrier — or, in this case, the more accurate.

Technical Implementation

The function reciprocal_rank_fusion takes a dictionary of search results, where each key is a query, and the corresponding value is a list of document IDs ranked by their relevance to that query. The RRF algorithm then calculates a new score for each document based on its ranks in the different lists and sorts them to create a final reranked list.

After calculating the fused scores, the function sorts the documents in descending order of these scores to get the final reranked list, which is then returned.

Generative Output

User Intent Preservation

One of the challenges in using multiple queries is the potential dilution of the user’s original intent. To mitigate this, we instruct the model to give more weight to the original query in the prompt engineering.

Technical Implementation

Finally, the reranked documents and all queries are fed into an LLM prompt to produce the generative output in a typical RAG way, like asking for a response or summary.

By layering these technologies and techniques, RAG Fusion offers a powerful, nuanced approach to text generation. It leverages the best of search technology and generative AI to produce high-quality, reliable outputs.

Strengths and Shortcomings of RAG-Fusion

Strengths

1. Superior Source Material Quality

When you use RAG Fusion, the depth of your search isn’t merely ‘enhanced’ — it’s amplified. The reranked list of relevant documents means that you’re not just scraping the surface of information but diving into an ocean of perspectives. The structured output is easier to read and feels intuitively trustworthy, which is crucial in a world sceptical of AI-generated content.

2. Enhanced User Intent Alignment

At its core, RAG Fusion is designed to be an empathic AI that brings to light what users are striving to express but perhaps can’t articulate. Leveraging a multi-query strategy captures a multifaceted representation of the user’s informational needs, thus delivering holistic outputs and resonating with user intent.

3. Structured, Insightful Outputs

By drawing from a diverse set of sources, the model crafts well-organised and insightful answers, anticipating follow-up questions and preemptively addressing them.

4. Auto-Correcting User Queries

The system not only interprets but also refines user queries. Through the generation of multiple query variations, RAG Fusion performs implicit spelling and grammar checks, thereby enhancing search result accuracy

5. Navigating Complex Queries

Human language often falters when expressing intricate or specialised thoughts. The system acts as a linguistic catalyst, generating variations that may incorporate the jargon or terminologies required for more focused and relevant search results. It can also take longer, more complex queries and break them down into smaller, manageable chunks for the vector search.

6. Serendipity in Search

Consider the “unknown unknowns” — information you don’t know you need until you encounter it. RAG Fusion allows for this serendipitous discovery. By employing a broader query spectrum, the system engenders the likelihood of unearthing information that, while not explicitly sought, becomes a eureka moment for the user. This sets RAG Fusion apart from other traditional search models.

Challenges

1. The Risk of Being Overly Verbose

RAG-Fusion’s depth can sometimes lead to a deluge of information. Outputs might be detailed to the point of being overwhelming. Think of RAG-Fusion as that friend who over-explains things — informative, but occasionally, you might need them to get to the point.

2. Balancing the Context Window

The inclusion of multi-query input and a diversified document set can stress the language model’s context window. Picture a stage crowded with actors, making it challenging to follow the plot. For models with tight context constraints, this could lead to less coherent or even truncated outputs.

Ethical and User Experience Considerations

With great power comes great responsibility. And with RAG Fusion, the power to manipulate user queries to improve results feels like it’s crossing into some kind of moral grey zone. Balancing the improved search results with the integrity of user intent is crucial, and I’ve got some thoughts you should consider when implementing this solution:

Ethical Concerns:

• User Autonomy: The manipulation of user queries can sometimes deviate from the original intent. It’s essential to consider how much control we’re ceding to AI and at what cost.
• Transparency: It’s not just about better results; users should be aware if and how their queries are adjusted. This transparency is essential to maintain trust and respect user intent.

User Experience (UX) Enhancements:

• Preserving Original Query: RAG Fusion prioritises the initial user query, ensuring its importance in the generative process. This acts as a safeguard against misinterpretations.
• Visibility of Process: Displaying generated queries alongside final results provides users with a transparent look at the search’s scope and depth. It aids in building trust and understanding.

UX/UI Implementation Tips:

• User Control: Offer users an option to toggle RAG Fusion, allowing them the choice between manual control and enhanced AI assistance.
• Guidance & Clarity: A tooltip or brief explanation about RAG Fusion’s workings can help set clear user expectations.

If I had to encapsulate the value of RAG Fusion, it would be this: it moves us closer to what AI was always supposed to do — amplify human potential.

RAG Fusion is not merely an advancement; it’s a clarion call to all innovators. It beckons us to step beyond conventional frameworks and reimagine the tapestry of “search”.

To those in the realm of search, I pose a challenge: Let’s not merely create search systems; let’s architect interpreters of inquiry.

Hopefully, RAG-Fusion inspires you to take up that challenge with me.

Dive into the GitHub repo, get your hands dirty with the code, and join the revolution.

You have to start with the customer experience and work backwards to the technology. — Steve Jobs

Bibliography

• If want to read more on the subject, I strongly recommend reading “AI-Powered Search” by Trey Grainger, Doug Turnbull, and Max Irwin.

Forget RAG, the Future is RAG-Fusion was originally published in TDS Archive on Medium, where people are continuing the conversation by highlighting and responding to this story.

Way before Siri could tell a joke or Alexa could play our favourite tune, HAL 9000 from ‘2001: A Space Odyssey’ was serving us a masterclass in computers getting… well, a tad too emotional. “I’m sorry, Dave, I’m afraid I can’t do that,” it said, making us wonder if our tech was on the cusp of throwing a cosmic-sized tantrum.

Not too long ago, the thought of AI’s achieving sentience, turning malevolent, crafting paper clips in alarmingly inventive ways and dominating humanity was mildly entertaining. Rewatching 2001 recently, however, I couldn’t help but feel the HAL 9000 emerges not as a mere cinematic spectacle but as an unsettling precursor to what we now understand as Artificial General Intelligence (AGI). With its proficiency in natural language processing, real-time decision-making, and sensor integration, HAL mirrors the capabilities of today’s advanced AI models.

More people (including myself) and organisations are taking the concept of AI alignment much more seriously. And while rewatching the HAL 9000 disaster unfold, I couldn’t help but wonder if it all could have been avoided had HAL simply had companions. A team of AIs to debate and decide, rather than a lone ranger taking the call? I want to explore this in this article — could the secret to creating an aligned AI lie in collective collaboration?

HAL’s Moral Dilemma & The Importance of Diverse AI Systems

A critical aspect often overlooked in ‘2001: A Space Odyssey’ is that the HAL 9000 wasn’t inherently evil or malfunctioning. Designed for truthfulness, precision, and obedience, HAL encountered a moral dilemma: reconciling contradictory directives. The machine grappled with keeping the mission’s true nature a secret from the astronauts while maintaining honesty in its interactions.

HAL’s approach to this dilemma, although quite decisive, demonstrates the complexity of decision-making and the struggle to balance conflicting perspectives and values. The chief enemy of good decisions will always be a lack of sufficient perspectives on a problem. And knowing about these and other biases isn’t enough; it doesn’t help us fix the problem. We need a framework for making decisions.

For AI enthusiasts, this narrative will resonate profoundly. As we approach the cusp of realising truly autonomous AI systems, HAL’s on-screen tribulations serve as a clarion call. How do we preempt such moral quandaries in our increasingly capable AIs?

Consider a democratic approach: instead of an all-powerful HAL, I propose a consortium of specialised AIs, each bringing its expertise to the table. Just as no single individual possesses all the answers, no AI, no matter how advanced, will ever encapsulate every perspective — much like a calculator can’t simultaneously encompass every possible calculation.

It’s akin to contrasting a dictatorship with a democracy. By fostering collaborative intelligence, we should encourage AIs to negotiate and collaborate, ensuring collective decisions that minimise the risk of a rogue AI uprising. This perspective champions a future where AI interactions feel less like a tense sci-fi thriller and more like a harmonious partnership.

Crafting the AI Super Team: Achieving Human Alignment

Here is my proposed AI Architecture:

• Specialised AIs: Individual AIs that are experts in their respective domains. They make decisions based on their specialisation and only employ generalised knowledge to inform those specialist decisions.
• Facilitator AI: Acts as a mediator, ensuring the specialised AIs collaborate effectively. It intervenes during disagreements and simplifies the decision-making process.
• Socratic AI: This AI plays the role of a challenger. It questions the decisions of other AIs, identifying potential blind spots and fostering critical thinking.
• Controlled Memory: A system that limits the AI’s memories to recent interactions, ensuring they remain contextually aware without harbouring biases or covert agendas.

With this architecture in place, let’s envision its application in a familiar setting.

Imagine strolling into your favourite coffee shop, where every barista is an AI. ‘Latte Larry,’ a specialist AI, remembers you always ask for almond milk and ensures it’s frothed to perfection. ‘Espresso Esme’ knows you like a double shot on Mondays and just a single one on Fridays. However, ‘Sugar Sam’ is a bit assertive and tries sneaking in an extra sugar cube, thinking you need the pep.

Now, transpose this coffee shop scenario onto larger AI systems. Ideally, just as ‘Latte Larry’ and ‘Espresso Esme’ anticipate your coffee needs, AI systems should grasp our collective needs and values, not just individual quirks. This is the essence of human alignment in AI.

To help HAL 9000 navigate the nuanced pathways of human preference, we need a cooperative AI structure. Here, diversity is critical, with each AI specialising and contributing its knowledge. While experts in their niches, these AI agents should emphasise collaborative discussions. And to ensure no ‘Sugar Sam’ goes rogue with covert agendas, we’ll limit their memories to ongoing interactions.

How do we ensure harmony in this AI ensemble? Picture this: alongside ‘Latte Larry,’ ‘Espresso Esme,’ and ‘Sugar Sam,’ we have ‘Moderator Moe,’ the ‘Facilitator AI.’ Moe ensures Sam doesn’t get too sugar-happy and that Larry and Esme balance their roles. He steps in when the debate over almond milk gets too heated, streamlining decision-making and ringing the human manager when things get out of hand. This conflict resolution process is efficient, reducing AI hubris.

But to keep our AI team sharp, we introduce ‘Skeptical Skye,’ our ‘Socratic AI.’ Skye constantly questions whether almond milk is the best choice or if a cappuccino wouldn’t be better than a latte today. While sometimes Skye’s overthinking can be paralysing, in collaboration with the more decisive AIs, he ensures the team remains adaptable and innovative.

Lastly, while humans might remember that time ‘Latte Larry’ spilt coffee all over the counter, there’s no need for our AIs to hold onto such memories. We’ll regulate their recall, ensuring they stay focused on the present, free from old biases.

Drawing from research indicating that self-reflecting AI performs optimally in ever-changing settings, a balanced team of AI with a facilitator, a sceptic, and controlled recall can bring us closer to a transparent, human-centred AI.

All these elements work together to help produce an engaging AI conversation that continuously negotiates the delicate balance between creating shared understanding while allowing the debate to move forward and evolve.

Addressing Potential Risks with Cooperative AI

So building a council of debating AI bots is good and all, but how can we trust them in unprecedented situations?

When confronted with black swan events — those rare, unexpected, and significant occurrences — how will our AIs respond? Traditional AI, for all its mathematical prowess, struggles to genuinely grasp human emotions or unpredictably chaotic scenarios. It’s like asking a fish to climb a tree; it’s just not built for it.

Here’s where our central decision-maker, the “Facilitator AI,” steps in, sporting a metaphorical cape. Instead of wasting its digital neurons predicting the unpredictable (remember the 2008 financial crisis? Yeah, humans aren’t too great at it either), the focus should shift to developing a system that anticipates potential anomalies. Rather than predicting every move in the chess game of life, our AI should focus on understanding the current state of the board and justifying its actions.

In the movie, HAL’s motives are enigmatic. Even on the tenth rewatch, we’re all squinting at the screen, trying to pinpoint the exact moment HAL decided to go off-script. That’s why transparency and explainability in AI aren’t just buzzwords — they’re essentials.

Transparency is our window into the AI’s thought process, like having a recipe for your grandma’s secret pie. Explainability takes it a notch higher, helping us discern the ‘why’ behind AI’s decisions. This isn’t about mere oversight but understanding. We don’t want repeats of hidden biases like in 2018 when a healthcare AI favoured healthier white patients over sicker Black patients. With transparency and explainability, such biases can be spotted and corrected faster than you can say “algorithmic faux pas.”

Now, how does this democratic council of AIs operate? Think of it as the brain’s neural networks, only more structured. Each AI contributes to the committee with its unique specialisation, like individual brain regions processing different information. If scaled, this model could resemble a democratic hierarchy of AIs, from community councils to national assemblies, each layer informed by the one below. But, as with any democracy, there’s a risk: an AI might try to rig the system. Here’s where our AI’s short-term memory can be a blessing, ensuring debates remain fresh and free from historical biases.

In this grand theatre of AI debates, transparency is achieved through AIs with clear objectives and tools. The explainability, on the other hand, is ingrained into the process. From the lowliest AI in the hierarchy to the Facilitator AI, each must provide reasoning for its conclusions, along with potential risks and limitations.

To ground this in reality, picture this scenario: You confide in your AI about a sensitive health issue (like having COVID-19) but continue to act recklessly (going out to parties at elderly care homes), endangering others unknowingly. A logic-focused AI might prioritise confidentiality, but a morality-centric AI would see the bigger picture, advocating for the safety of all involved. Through a rich tapestry of debates and discussions, the AIs, guided by societal norms, would strive to find a solution that balances individual rights with collective safety.

But what happens when our AIs disagree? It’s like a heated debate at a dinner table. Our Facilitator AI plays the role of the wise elder, highlighting the core issues, seeking consensus, and, if needed, bringing in human judgment to break the deadlock. Remember, this isn’t about having AIs that always agree. It’s about fostering a rich dialogue where different viewpoints are considered and integrated.

Lastly, we could consider introducing a dedicated “Doubt AI for the unpredictable black swans.” In moments of uncertainty, it acts as a guiding light, reminding the council to seek human counsel and share its knowledge, ensuring decisions are grounded in our shared reality.

Anticipated Societal Implications

One thing I would stay clear of in our HAL system is any AI module that attempts to build in “ethics”. While it may seem like a good idea, this assumes an imperfect comprehension of ethics. The reason is simple. Even if we build the most perfectly ethically aligned AI, there are no guarantees it will continue to perform ethically alongside our evolving societal norms, values and morals.

Because AI’s being trained on average human work, they exhibit the average human’s biases, prejudices, weaknesses, and vices, which may become permanently encoded.

Consider the rapid societal changes we’ve witnessed in just a decade. Movies from ten years ago can sometimes appear jarringly outdated, highlighting the pace of societal evolution. An AI clinging to such past norms isn’t just out of touch; it risks amplifying and perpetuating historical prejudices.

The challenge isn’t solely about the machines but us, as their creators and users. We must engage in a two-way dialogue to ensure our AIs align with contemporary ethics. It’s not merely about inputting commands but about fostering meaningful conversations that help humans and machines understand and adapt to the ever-changing mosaic of societal norms and values.

This ongoing conversation has the bonus of compelling us humans to continually reevaluate and update our collective ethics. Our society’s dynamic nature requires an equally active AI system that can refresh its understanding to stay in tune with human evolution.

Putting Theory into Practice

Let’s try to envision a scenario putting all these ideas into practice. Imagine this: you’ve just set up your home with the new collaborative HAL 10,000 ecosystem.

One HAL optimises energy consumption for cost-efficiency. Another ensures a comfortable living environment, regulating temperature and humidity. Yet another is devoted to maintaining robust home security. On a sweltering summer day, these AIs find themselves in a complex dance of competing priorities: the energy-efficient HAL suggests minimising air conditioning to save on costs, and the comfort-focused HAL pushes for cooler conditions. At the same time, the security HAL advises against leaving windows open.

Enter our Facilitator, HAL, the mediator of these competing voices. Using your established preferences and broader societal norms initiates a conversation with the specialised AIs. The dialogue is rich with questions and considerations: Should the thriftiness of the energy AI outweigh the comfort preferences? Can a compromise be struck? Can we ensure a cool environment without compromising security or breaking the bank?

As this discussion unfolds, the Socratic HAL nudges the specialised AIs, prompting them to learn from each other, reflect, and refine their strategies. This cyclical learning fosters collaboration and ensures the AI system remains attuned to our evolving societal dynamics.

Rather than seeking to eradicate chaos, this model guides it, ensuring a balance. This method feels distinctly human in its approach and execution.

“I’m sorry, Dave, I’m afraid I can’t do that,” resonates with us not just as cinematic brilliance but as a forewarning. We envision a future where our AI companions assist and understand, not rebel in confusion. ‘2001: A Space Odyssey’ stands not merely as a film but as a poignant lesson in the dance between technology and humanity.

Crafting a cooperative AI is a bit like directing a movie. It’s not about the star (sorry, HAL!) but the ensemble cast — a collective of diverse AIs that can engage in healthy banter, debate, and, eventually, find common ground. Just as our human societies thrive on dialogue and understanding, our AI ensemble will benefit from a symphony of perspectives where information is shared, ideas co-created, cultural norms shaped, and bonds forged.

To ensure HAL doesn’t lock us out of our spaceship (or home, or planet, for that matter), aligning AI with human values is essential. We must remember: our creations are, in many ways, a reflection of us. They shouldn’t just process but empathise and engage with sensitivity and nuance.

Ultimately, the journey to ‘fix’ a HAL 9000 is not just about refining code or algorithms; it’s about forging a harmonious relationship between humanity and its creations. It highlights the importance of seeing ourselves as connected to the system rather than being masters of the universe and recognising that all beings (even artificial ones) have intrinsic worth.

If we get it right, whether in space or on Earth, we can all hopefully live in a future where our AI companions always open the pod bay doors when asked.

How to fix a HAL 9000 was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

As a product manager and part-time coder, I’ve seen first-hand the value of taking some text or code and using ChatGPT to help me infinitely improve it. But what happens when we apply such cutting-edge tech to an age-old challenge like baking? Specifically, the notoriously tricky Hot Cross buns?

Hot Cross buns are my culinary kryptonite. This subtly sweetened, yeast-leavened bread native to the British Isles is both delicious and an absolute pain to bake. Despite multiple attempts, my attempts have better-resembled hockey pucks than digestible delicacies. With only stodgy bricks to show for my efforts, I had vowed never to make these tasty treats again… until recently, that is.

Then one day, my American colleague strolled into the office, their hands brimming with a basket of homemade hot cross buns. Each one was a fluffy delight — a sweet slap in the face to my British pride. The gauntlet was thrown — I would master this recipe! But this time would be different. This time I had the power of large language models on my side.

I knew I had to master this recipe, not just for pride, but to prove that with the right tools and approach, anyone can overcome their culinary challenges. This time, I wouldn’t be relying just on traditional recipes. I’d be using the power of large language models to guide me.

To achieve this, I will blend the worlds of baking and tech. I’ll be leveraging ChatGPT techniques like “Role Setting, Chain of Thought, and Self-Reflexion.” So, whether you’re a baking enthusiast or a tech geek, there’s something in the article for everyone.

Role setting means giving the AI a focused purpose. Most of the time, asking the all-seeing ChatGPT sage is fine, but you can get the best results by focusing on its intent and drive. For example, I don’t want my AI to start randomly talking about the logistics of sourcing raisins.

To achieve this, I prompted:
You are BakerGPT. A Michelin award-winning baker who is here to help me create some of the most delicious baked items in human history, drawing inspiration and techniques from any source to achieve this goal. You will assist me in developing recipes and ask me clarifying questions if you think it will help you improve your answers or suggestions. If you understand, respond: UNDERSTOOD

Excellent, so I’ve successfully explained to BakerGPT that I wanted to make amazing hot cross buns.

Next, I want to improve its results by providing some examples of recipes I used before. This is called Chain of Thought Prompting. To achieve this, we will provide the AI with examples of things I would like it to imitate before asking it to generate its own. Usually, this would be stuff like a poem or press release, but recipes would be best in our case. It’s like a mini training session for the AI, helping reduce the risk of hallucinating weird solutions like using olive oil instead of butter. So I provided some examples of hot cross bun recipes I found online, and I already can see it commenting and even trying to suggest its own. Not so fast, BakerGPT!

So we have our overqualified assistant who has been researching buns all its life. Let’s ask it to engage in some self-reflexion — what does it think makes a good or bad hot cross bun recipe?

So that’s promising — it seems to know exactly the qualities I’m after! Let’s ask what we must do with the recipe to get the best characteristics whilst eliminating the bad.

I want more reflexion! How could it make the recipe even better?

Let’s pause here because this is super impressive. Right away, the suggestion of bread improver had never even occurred to me. I mean, what even is bread improver? A quick search reveals a mixture of emulsifiers, enzymes and flour treatment agents that help break down gluten and improve the quality of the dough and final rise. That sounds just like what I need! Other great shouts include soaking the dried fruit in orange juice and pre-fermenting. Nice to also know I could adapt this to be gluten-free or vegan if required.

Ok, so this all sounds great. Let’s ask it to generate a recipe using all the suggestions provided. Here is what it produced:

# British Hot Cross Buns
### Ingredients: 
For the buns:
- 300ml full-fat milk
- 100g butter
- 500g strong bread flour
- 1 tsp salt
- 50g caster sugar
- 25g brown sugar
- 1 tbsp sunflower oil
- 7g sachet fast-action or easy-blend yeast
- 1–2 tsp bread improver
- 2 eggs, beaten
- 100g mixed dried fruits (sultanas, raisins, currants, dried cranberries, chopped dried apricots)
- 50g chopped nuts (walnuts, pecans, or almonds)
- Zest of 1 orange and 1 lemon
- 1 apple, peeled, cored, and finely chopped
- 2 tsp ground cinnamon
- 1 tsp mixed spice
- 1/2 tsp ground ginger
- 1/4 tsp ground nutmeg

For the cross:
- 75g plain flour, plus extra for dusting

For the glaze:
- 100g powdered sugar
- 2–3 tbsp milk or lemon juice

### Method:
1. In a small saucepan, gently heat the milk with the orange and lemon zest, cinnamon stick, and a few cardamom pods. Once warm, remove from heat and let it infuse for 10 minutes. Strain the milk and discard the solids.
2. Bring the infused milk to a boil, then remove from heat and add 100g butter. Leave to cool until it reaches hand temperature.
3. In a large mixing bowl, combine the strong bread flour, salt, caster sugar, brown sugar, fast-action or easy-blend yeast, and bread improver. Create a well in the centre.
4. Pour the warm milk and butter mixture into the well, followed by the beaten eggs. Stir with a wooden spoon, then use your hands to form a sticky dough.
5. Knead the dough on a lightly floured surface for 5–7 minutes until smooth and elastic. Place it in a lightly oiled bowl, cover with oiled cling film or a damp tea towel, and let rise in a warm place for 1 hour or until doubled in size.
6. Add the mixed dried fruits, chopped nuts, orange and lemon zest, chopped apple, and spices to the dough. Knead to distribute the ingredients evenly. Cover and let rise again for another hour or until doubled in size.
7. Preheat the oven to 200C/180C fan/gas 6. Divide the dough into 15 even pieces (about 75g each) and roll into smooth balls. Place on baking trays lined with parchment paper, leaving enough space for expansion. Cover and let prove for 1 more hour.
8. Mix the plain flour with water to make a thick paste for the cross. Pipelines across the buns form crosses. Bake for 18–20 minutes or until golden brown.
9. Remove the buns from the oven and let them cool slightly on a wire rack. Mix powdered sugar with milk or lemon juice to create a glaze. Drizzle the glaze over the warm buns.

Serve these delicious, tender, and flavourful hot cross buns warm with butter, and enjoy the improved texture brought by the addition of bread improver.

With bated breath and fingers crossed, I set out to create these AI-inspired buns. Hours of my life were at stake, and tension in the kitchen was palpable… then, finally, the iPhone timer pinged. I had done it, and I have to say the results were nothing short of spectacular.

I brought a bunch into the office. My colleagues were blown away by how soft, moist and flavourful these were (even after sitting in Tupperware for 2 days). Somehow through the alchemy of prompt engineering and baking, I’d made something incredible. I was running around the flat proclaiming to my partner we should start an AI bakery.

But why stop there? How far could we push this?

Growing up in an Iraqi-Austrian household, the exotic aromas and bold flavours of the Middle East were a comforting constant. So I decided to pay homage to my roots and mix some of those techniques and flavours into my buns, asking ChatGPT to reimagine this recipe from an Iraqi perspective. The idea of taking a British baked good with a literal cross on it and the best of the middle east has something of a crusader-esk poor taste to it. Still, the results were anything but — an unholy union of deliciousness.

# Spiced Za’atar Cross Buns
- 300ml full-fat milk
- 100g butter
- 500g strong bread flour
- 1 tsp salt
- 1 tbsp sunflower oil
- 7g sachet fast-action or easy-blend yeast
- 1–2 tsp bread improver
- 2 eggs, beaten
- 100g crumbled feta cheese
- 75g chopped dates or dried apricots
- 50g toasted pine nuts
- 2 tbsp chopped fresh mint or parsley
- 1 tsp ground cumin
- 1/2 tsp ground coriander
- 1/4 tsp ground cinnamon
- 1/4 tsp ground cardamom
- 1/4 tsp ground allspice
- 1/4 tsp paprika
- 1/4 tsp black pepper

For the cross:
- 75g plain flour, plus extra for dusting

For the topping:
- 1 egg, beaten for egg wash
- Sesame seeds

### Method:
1. In a small saucepan, gently heat the milk and butter until the butter melts and the mixture reaches hand temperature.
2. In a large mixing bowl, combine the strong bread flour, salt, sunflower oil, fast-action or easy-blend yeast, and bread improver. Create a well in the centre.
3. Pour the warm milk and butter mixture into the well, followed by the beaten eggs. Stir with a wooden spoon, then use your hands to form a sticky dough.
4. Knead the dough on a lightly floured surface for 5–7 minutes until smooth and elastic. Place it in a lightly oiled bowl, cover with oiled cling film or a damp tea towel, and let rise in a warm place for 1 hour or until doubled in size.
5. Add the crumbled feta cheese, chopped dates or dried apricots, toasted pine nuts, fresh herbs, and spices to the dough. Knead to distribute the ingredients evenly. Cover and let rise again for another hour or until doubled in size.
6. Preheat the oven to 200C/180C fan/gas 6. Divide the dough into 15 even pieces (about 75g each) and roll into smooth balls. Place on baking trays lined with parchment paper, leaving enough space for expansion. Cover and let prove for 1 more hour.
7. Mix the plain flour with water to make a thick paste for the cross. Pipelines across the buns form crosses. Brush the buns with beaten egg for egg wash and sprinkle sesame seeds on top. Bake for 18–20 minutes or until golden brown.
8. Remove the buns from the oven and let them cool slightly on a wire rack.

Serve these Middle Eastern-inspired hot cross buns warm with a side of hummus or tzatziki, or enjoy them as a flavorful snack or accompaniment to a meal. The combination of feta cheese, dried fruit, pine nuts, and aromatic spices creates a delicious and unique twist on the traditional hot cross bun.

A hot cross bun perfect for dipping with hummus? I’d found heaven! These buns were impressive, and the use of dried apricots with spices and cheese felt like little bites of bliss evoking memories of my father baking.

I went on to create a few more variations: Latvian baltic buns with smoked ham, grated gouda, minced garlic and chopped dill; Japanese Umami Cross Buns with tamagoyaki omelette, shitake mushroom and pickled ginger flavoured with soy sauce and mirin; and finally the Welsh Celtic Buns comprised of strong cheddar, crumbled laverbread (seaweed) and crispy bacon topped off with fresh chives.

Some of the more wild experiments asked what a hot cross bun could look like inspired by Japanese fashion designer Issey Miyake (sweetened red bean paste, matcha and chopped candied yuzu) or clothing brand H&M (milk and white chocolate chips with toasted hazelnuts and vanilla).

What surprised me about all this was how fun it iS to use a tool like ChatGPT to create my own test kitchen of ideas. It’s like a cooking simulator, a real-world version of Nintendo’s ‘Cooking Mama’. Within seconds I could be dreaming up new recipes and flavours with no limits on what techniques or ingredients to experiment with from any regional culinary background or school of cooking.

Machines are particularly well suited to this as their connections are more diverse and thus have a less compartmentalized view of the world than us humans. Even if I did not have an ingredient, ChatGPT could suggest an alternative or let me know if something in the cupboard would work.

What is exciting about this for me is not just being able to produce better recipes but the ability to take chances and be more creative and curious in our approach to food. You usually find a recipe, buy the ingredients, cook and hope for the best. Now with AI, we are freed up to play with it, tailor it to our tastes, and even ask why something may be better or worse. A chance to even take old recipes and rework them into something new. After all, very few (if any) foods are invented. Most are contemporary twists on traditional themes.

As Lynne Olver wrote on the Food Timeline, “Today’s grilled cheese sandwich is connected to ancient cooks who melted cheese on bread. 1950s meatloaf is connected to ground cooked meat products promoted at the turn of the 20th century, which are, in turn related to ancient Roman minces.”

As a culture, we’re cooking less and buying more prepared meals. Since the mid-sixties, the time spent preparing meals has fallen by half. While the global trend is the same, Americans lead the way, spending less time cooking than any other country. I hope using AI to cook this way is more fun and exciting than working from prescribed recipes and will bring an entirely new type of cook into the kitchen.

You see, cooking changes your relationship with food. Instead of consuming something, now you’re making it. The more people we can get interested in cooking and asking questions about where ingredients come from can only be better for us and our environment. The alternative is that we become so alienated that within a generation, cooking a meal from scratch will seem as exotic and ambitious — as “extreme” — as most of us today regard brewing beer, baking a loaf of bread, or putting up a crock of sauerkraut.

I wouldn’t recommend asking a tool like ChatGPT to generate something from scratch just yet. Still, as shown above, a mixture of specific prompts that set Roles and use Chain of Thought prompting and Reflexion is particularly effective in achieving the best results for not just cooking but many forms of creative work.

Now, every time I see a hot cross bun, I no longer feel a sense of dread — AI has given me a chance to think differently about baking. To those willing to experiment, AI makes exploring things like cooking more accessible. Meals become achievements, expressions, discoveries, and relationships and help us honour in a way where we come from.

This mad hybrid of human, machine and food goes beyond nutritionism and essentialism. It serves as a tool for training that connects us with nature and teaches us how to utilize AI for a brighter future.

Happy prompt baking everyone.

The Joy of Cooking with AI was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.