Think

In the last post I showed some of the things twain does quietly in the background. This is the technical deep-dive and architecture underneath.

The hardest part of building Twain with AI wasn’t getting the model to say something. It was deciding when it shouldn’t.

> twain uses a local-first architecture: Deterministic code for things that should simply be correct, and a language model where judgement is actually useful.

Drawing that line shaped almost everything.

Search, recurrence, date grouping, charts and calculations happen locally on the phone (also privacy reasons). Semantic vectors use Apple’s NLEmbedding. Retrieval is hybrid: exact words, shared people and themes, plus semantic similarity.

The model handles the things code is bad at: reading handwriting, deciding whether two entries genuinely connect, choosing a line worth keeping, or writing prose someone will actually read.

Charts and calculations happen in code. The model only gets the resulting facts and uses them to write the human-readable interpretation. I’m using AI for judgement, not arithmetic.

The same principle applies to model output.

> twain uses structured outputs - typed JSON against a schema - rather than repeatedly asking the model to “chat about” the journal. Anything presented as your own words is then checked deterministically.

A quote has to exist verbatim in the source.
A tag has to appear in the entry.
An emphasised phrase has to be a real substring.

The model proposes. Code gets to refuse. Several prompts are explicitly designed to return null.

Most days do not contain a profound insight. Two entries mentioning work and a walk are not necessarily connected. If there is nothing worth showing, the correct output is...nothing.

> Another decision I’ve come to love: derived AI output is disposable.

Reflections, digests, stories and readings are cached against a hash of their source material. Change an entry and they can be regenerated. Interrupted work resumes. Missing weeks and months backfill. Sixteen self-healing routines repair derived state rather than asking the user to fix it.

> And the journal itself stays local.

13 SwiftData models live on-device. There is no account, sync layer or cloud database holding journal content. The only server component is a 195-line relay that holds the API key and usage limits; it does not parse, store or log the journal.

In architectural terms: local-first storage, hybrid retrieval, structured outputs, deterministic guardrails, and idempotent, resumable pipelines.

From first commit to this version: 27 days, 71 Swift files, 17,212 lines and 116 commits.

> The product principle underneath all of it is simpler:

Don’t ask AI to do everything.
Use it where judgement helps. Use code where correctness matters. And make the model earn the right to show something about a person’s life.

I've written up a full deep-dive how twain is built here for those that want to learn more: https://lnkd.in/esE343Ea

One of my favourite things about twain is how much it does without you ever asking it to.

I never wanted to build a journal app where the “AI feature” is just a chat box.

The ambition was different: What if the journal itself gradually became more intelligent the longer you used it?

A lot of twain therefore happens quietly in the background.

A few of my favourite examples:

> It learns your handwriting.
Correct a misread word once and Twain remembers it. Your vocabulary and recurring handwriting mistakes feed forward, so transcription gets better the more you use it.

> It remembers meaning, not just words.
Every page indexes itself for exact and semantic search, so you can find something months later even if you no longer remember how you phrased it.

> It figures out who people are.
If “Max” starts looking like two different people, Twain asks. Define them once and it can separate those threads across your journal rather than quietly mixing two lives together.

> It finds echoes across years.
New entries are compared with the past using people, themes, words and semantic similarity. Older strong matches are deliberately given a chance too, otherwise “memory” quickly becomes “things you also wrote about last week”.

> It notices before you ask.
That retrieval powers Twain noticed: sometimes a thought about today, sometimes a connection to something you wrote years ago.

> It brings old chapters back at the right moment.
On the exact anniversary of an entry, Twain can compare then with now and create a Story when there is something genuinely worth resurfacing.

> It knows when not to.
If there is no meaningful connection, no Story appears. If there is no line worth keeping, it doesn’t invent one. Sometimes the smartest thing it can do is leave you alone.

> It quietly maintains itself.
Weeks and months compose in the background. Correct an old entry and anything derived from it can regenerate. Interrupted work resumes. Scan a stack of old notebooks and the archive gradually catches itself up.

And all of this happens without turning your journal into an analytics dashboard.

No streak pressure. No “top five topics”. No arbitrary score telling you whether your week was good or bad.

The goal is something that feels less like software you operate and more like a companion that has actually been paying attention.

And importantly, the journal itself stays on your phone. No cloud database. No remote copy of your writing. Privacy as a feature, not an after-thought.

There are still plenty of things I want twain to get better at.

But this is probably the product idea underneath all of it that excites me most:

AI not as a feature you open, but as intelligence woven quietly through the product.

This is probably the most selfish product I’ve ever built. Because it’s the product I’ve always wanted for myself.

In my last post a few days ago (which led to a lot of lovely comments and messages, thank you for that btw) I mentioned one of the things I’d spent the past few weeks building.

A journal app from an idea I’ve carried around for years.

This is it.

It’s called twain.

I’ve been journaling for many years and written close to 200 pages by hand this year alone.

There’s something about putting pen to paper that I don’t think a screen or app can quite replicate. And I never wanted twain to replace that.

What I’ve always wished for instead was a way to unlock everything that accumulates inside those notebooks over time.

Because you write your life down without really knowing what will matter later.
A passing thought. A person you’ve just met. Something you’re scared of. The aftermath of a heartbreak. The excitement of a new job. A completely ordinary Tuesday. The first signs of a chapter beginning, or ending, before you even realise it.

And then you close the notebook.

Years of your life can end up sitting on a shelf, full of people, places, feelings, questions, patterns and versions of yourself you’ve almost forgotten.

There is so much of who we are hidden inside those pages.

That’s what I wanted twain to make visible.

Not by replacing the journal. The opposite.

You keep writing with pen and paper, exactly as you always have. twain sits alongside it as a companion: preserving the pages, making them searchable, noticing the threads that run between them, and occasionally bringing something back from the past that suddenly means something again in the present.

The notebook stays the hero. And it deserves to be remembered.

Handwriting is imperfect. Journaling is imperfect. That’s part of the point.

There are crossed-out words, unfinished thoughts, terrible handwriting, contradictions and things that only made sense to you on that particular day.

That’s part of what makes it honest. twain isn’t trying to correct that. It’s trying to preserve it.

Somewhere along the way, twain has also grown into far more than the little journal companion I initially imagined. There’s already much more inside it than I could sensibly cover here. And plenty more I want to build.

But more than anything, I’d love for people who genuinely journal to experience it and tell me what feels meaningful, what doesn’t, and what Twain should become.

It’s currently in a closed beta on TestFlight.

If you keep a handwritten journal and would like to try it, drop me a note or request access at thetwain.app.

Leave your mark (literally).

PS: Will share more about the different features and underlying architecture over the next days. It still blows my mind that I, without any CS background, managed to build this from the ground up. We live in special times.

Back in March, I decided to leave Healf. That chapter officially came to an end in June.

Increasingly, I felt a strong pull to get closer to building again, and a deeper desire to work at the forefront of two spaces I care deeply about: Health and AI. That made the decision relatively straightforward. Other factors played a part too, but they’re not important here.

There are many things I take away from the Healf chapter. Most importantly, a small but exceptional group of people I know I’ll build great things with again.

I stayed quiet afterwards. I wanted space without needing to attach a new title or perfectly formed plan to whatever came next.

So I’ve spent the last few weeks doing the things that always get pushed aside when you’re busy with all the “important” things at work:

Thinking without an agenda.
Reconnecting with people.
Following curiosity.
And, most importantly, being creative and building.

Earlier this year, I wrote that I’d never felt more enabled as a builder.

AI has collapsed the distance between imagining something and putting something real into the world.

Think it.
Design it.
Build it.
Ship it.

I’ve spent much of the summer doing exactly that.

I finally built a journal app from an idea I’ve carried for years. I started building an AI agent for myself to understand agent architecture, which has since taken on a life of its own. And alongside that, I’ve been working on a bigger idea in preventative health, where we’ve just started our first real-world pilot.

And somewhere in all of this, an older ambition keeps resurfacing: building something meaningful in mental health. I’m starting to follow that thread more seriously. Let’s see where it leads.

I’ll share each over the next few weeks.

A lot of the last two years of my life has been shaped by change - professionally and personally.

New roles. New people. New ideas. Endings I expected and some I didn’t.

Things I thought would last that didn’t, and opportunities I couldn’t have imagined beforehand.

No matter how rational or stoic you try to be, you don’t go through enough change completely unaffected.

Some changes are exciting. Some uncomfortable. Some take longer to make sense of than you’d like.

But they create space. And more often than not, something new finds its way into it.

None of what I’m doing today was part of a neatly constructed next-career plan.
I’ve mostly been following curiosity, learning, and most importantly my own intuition, making things I want to exist and seeing where they lead.

Some may become companies. Some may remain projects. Some paths are becoming clearer. Others are still unwritten.

As one chapter closes, another opens. Sometimes several at once.

It’s also a chance to prove that you can show up for yourself when nobody is watching.

I’ve rarely felt more energised, curious and excited about what’s ahead.

Building mode fully enabled.

the most interesting thing about a health score is not whether it is accurate. it is what it makes you believe, do, and feel next.

i recently came across a tweet from Aakash Gupta referencing a 2014 Colorado College sleep study that i haven’t been able to stop thinking about.

participants were given fake feedback about how much REM sleep they had supposedly gotten. the people told they had slept “better” then performed better on parts of a cognitive test battery.

in other words: the number changed the outcome.

that’s such an important idea for modern health-tech.

because it suggests that scores are not just passive mirrors of the body.
they can become interventions.

we have all be there:

you check your sleep score.
you see a bad number.
you subtly lower your expectations for the day.
you conserve effort.
you interpret normal variation as dysfunction.
and the score starts shaping the very state it claims to measure.

there a growing under current of users starting to be more aware of the psychological impact of "blind tracking". Ellie-mae Hammond nicely put it in her recent article "The Case For Breaking Up With Your Health Tracker" for British Vogue:

"My mornings no longer feel like my own. They belong instead to my Oura ring and my Garmin watch, harbingers of either joy or gloom depending on what the coloured circles and scores reveal about my health and general wellbeing once I wake. A green circle inspires relief, while amber prompts mild anxiety. A red circle can compromise my whole day."

that is where i think a lot of wearable products still fall short.

not because the data is useless.
and not because the sensors are bad.

but because data without context can be directionally wrong for the person living it, and can therefore become anxiety and obsession inducing.

a low score might reflect poor recovery.
or emotional stress.
or excitement.
or a disrupted routine.
or nothing especially meaningful at all.

and without subjective context, a personal baseline, and some sense of what is actually normal for you, the number can become louder than your own experience.

i was thinking about this again after seeing recent WHOOP study results shared by Gregory Grosicki, PhD showing a decline in self-reported alcohol consumption over time among a subset of their users.

that might reflect real behaviour change from feedback and self-monitoring.

but it is also a good reminder of how careful we need to be with causality in this space. people who track more are often already the people trying to change.

either way, i think the broader point is the same:

the job of health-tech is not just to measure people.

it is to help them interpret what a signal means for them, in the context of their actual life, without turning every score into identity.

you’re not a number.
but numbers can still change you.

that’s exactly why the personal, n=1 interpretation layer matters so much.

I’ve recently started playing around with Open Wearables, and honestly, this feels like a very good direction for the space.

not because it is another shiny health api.

but because it seems to understand something important:
→ the future of health products probably won’t be built on more closed black boxes.
→ it will be built on more open infrastructure, more transparent algorithms, and more composable intelligence layers.

Open Wearables being open source, self-hosted, MIT-licensed, and built around a unified wearable api already makes it interesting. but the more compelling part is the philosophy underneath it: open health scores, auditable logic, customizable thresholds, and an ai reasoning layer that tries to move beyond just reading back numbers.

that feels directionally right.

especially in a world where open models, open weights, and open protocols are starting to unlock whole new categories of products much faster than closed platforms would like.

in health, that matters even more.

because if you care about trust, nuance, safety, and building products that can actually adapt to different domains, populations, and contexts, then openness is not just a developer preference. it’s a strategic advantage.

also love the ambition of not stopping at raw data ingestion, but trying to bridge the gap from wearable data → health scores → reasoning → recommendations.

one api, open algorithms for things like sleep, recovery, strain and hrv, plus domain-specific coaching profiles, is a pretty powerful starting point for anyone building in this space.

big respect to Piotr Sędzik Jan Kamiński and the rest of the Momentum team behind it. they launched on Product Hunt today - go have a look and give them some love 🔼 . link in the comment.

the most interesting health products of the next few years may not come from who owns the device. they may come from whoever builds the best interpretation layer on top of open foundations such as Open Wearables. excited to be watching this space.

i asked an AI whether a home gym in the toilet was a good business idea.

it replied:
“Toilet Gym is stupid enough to spread and just plausible enough to exist.”

and then it gave me five surprisingly convincing reasons why it could work.

funny, yes.
but also a bit unsettling.

because that is increasingly the risk I see with how LLMs are being used in the workspace.

show them an idea you already want to believe in, and they will often help you believe in it more.

not always because the idea is good.
often because the model is very good at (and incentivised) making a direction feel coherent, expandable, strategic, even inevitable.

that is a dangerous thing.

especially in a moment where more and more people are trying to use AI to short-circuit the hardest part of creation:

judgement. critical thinking.

the difficult bit was never just generating options.
it was sitting with uncertainty.
challenging your own assumptions.
testing whether something is actually true, not just well-packaged.
feeling the difference between a good idea, an exciting idea, and an idea you are emotionally attached to.

AI is brilliant at helping you elaborate.
it is much less reliable as a substitute for taste, discernment, lived context, or real critical thought.

and the trap is subtle.

the more you ideate with it, the more it can start to nudge you down a path that feels validated simply because it keeps producing intelligent-sounding extensions of your premise.

you're pulled deeper into the conversation feeling like you have momentum.

sometimes what you actually have is a beautifully articulated bad idea.

i don’t think the answer is to reject AI.
i use it all the time.

but I’m becoming more deliberate about when I use it.

more often now, when something really matters, i try to do the opposite first:

i shut the laptop.
i take a piece of paper and a pen.
i sketch.
i sit with the tension a bit longer.
i talk to other humans.
i let disagreement, discomfort, and slowness do their job.

only once i feel like there is something real there do i bring AI in to expand, pressure-test, explore edge cases, or deepen the thinking.

AI can sharpen thought.
but it should not replace the formation of thought.

because the biggest risk in an AI-everything world is not that the models will think for us.

it’s that we will slowly stop noticing when we’ve stopped thinking properly ourselves.

*** Full AI response in the comments for entertainment purposes only - not business advice :)

if the confidence gap is where a lot of wellbeing journeys stall, then the next stage is not more optimisation.

it’s personal understanding.

for me, that’s the layer the category still struggles to define properly.

because personal understanding is not:
→ having more data
→ getting more alerts
→ chasing the perfect score
→ consuming more supplements
→ or being told what works for people "like you"

it’s something quieter than that.

it's knowing your own patterns well enough that life starts to feel less random.

→ it's realising that yesterday's "bad day" wasn't random, it was the quiet aftertaste of a conversation you never finished.
→ it's noticing your worst sleep doesn't follow your busiest days, it follows the ones where your mind never found an off switch.
→ it's learning to tell the difference between a body that needs rest and a mind that's been carrying too much for too long.
→ it's seeing - clearly, in your own data - that no supplement or routine can substitute for a life that's drifted away from what actually restores you.
→ it's opening the app less, because you've started hearing what your body was always trying to tell you.

no tracker sees that. no blood panel can measure that. no journaling app discovers that.

that, to me, is the real breakthrough.

not optimisation.
not information.
not even “personalisation” in the way most apps use the word.

understanding.

the kind that builds over time.
that remembers context.
that gets sharper as life changes.
that helps you make better decisions with a little more calm and a little less guesswork.

That’s the layer missing between tracking and real agency.

and i suspect it’s where the next generation of health products will be won.

not by helping people measure more.

by helping them understand themselves well enough to live better.

following on from my last few posts, i’ve been trying to map the journey i think a lot of people now go through in modern wellbeing.

from the trigger,
to the control phase,
to optimisation theatre,
to the interpretation gap,
to the personalisation illusion.

i tried to capture that thinking in a simple chart below.

and the stage i keep coming back to most is this one:

*the confidence gap*.

because i increasingly think one of the biggest unmet needs in health-tech is not more insight.

it’s confidence.

not confidence that the app is clever.
confidence that what it is telling you is actually true for you.

most products can give you:
→ a score
→ a trend
→ a correlation
→ a recommendation

but far fewer help you answer the more important question:

can i actually trust that this pattern is real in my own life?
did i sleep better because i stopped eating late?
or because work was less stressful this week?
did my energy improve because of the supplement?
or because i finally had two quieter evenings in a row?
is this a meaningful signal?
or just noise i’m projecting a story onto?

that uncertainty is everywhere.

and i think it explains why so many people get stuck below what i’d call the current *optimisation ceiling*.

→ we try more things.
→ track more variables.
→ collect more data.
→ follow more protocols.

but never quite build enough confidence to break through into real self-understanding.

that, to me, is the *confidence gap*.

and i suspect it is one of the main reasons so many wellbeing journeys feel
like motion without the necessary *escape velocity* to enter into the true *understanding layer*.

not because we lack effort.
not because we lack information.

because we still don’t know, with enough confidence:

what actually works for me?

real confidence comes from something much more specific:
→ change one thing
→ observe what happens
→ repeat enough times
→ start trusting the pattern

in other words: not just advice, but learning.

that feels like a much more honest job for health products.

not pretending to know you instantly.
not throwing generic recommendations at you.
but helping you build evidence about yourself over time - enough evidence to move beyond guesswork and into understanding.

to me, that’s where the real value starts.

not when an app tells you what usually works for people like you.
when it helps you become more certain about what actually works for you.

curious whether others working in health, product, or ai see the same pattern.

#heathtech #healthai #wearables #health #wellbeing

WHOOP's round is not just a WHOOP story. it’s a strong category signal.

$575m at a $10.1b valuation is one thing.
but the more interesting part is the operating proof underneath it:

→ 2.5m+ members
→ 103% bookings growth
→ $1.1b run rate
→ cash-flow positive
→ 8+ app opens per day on average (mental for consumer health)

that tells us something important.

the demand for personal health is real.
the engagement is real.
the investor appetite is real (esp. from institutional investors - but that is for anther post).

people clearly want a more continuous relationship with their health.

but i also think this highlights where the next layer of opportunity sits.

wearables like WHOOP are incredibly strong at capturing objective signals.
what still feels underbuilt is the layer that helps people connect those signals to the texture of their actual lives:

→ who they are
→ what tends to work for them
→ what patterns are real
→ what to change
→ what to ignore
→ and how to make better decisions over time

in other words: not just more biometric data, but better self-understanding.
that’s where this category gets really interesting.

both WHOOP and ŌURA have started to layer some subjective data on top of the physiological data they track. this is a super interesting signal.

who's listening?

most health apps don’t really personalise. they segment. that sounds like a small distinction. i don’t think it is.

after years in product, i’ve noticed the same pattern again and again:

an app asks for your age, gender, goals, activity level, maybe a few preferences during onboarding.

then it puts you into a bucket and serves content, recommendations, and “insights” based on that bucket.

→ some add rule-based branching.
→ some layer in shallow correlations from your data.
→ a few make the experience feel dynamic enough that the personalisation claim sounds believable.

but most of it is still the same basic thing:

population patterns with your name on top.

and users feel that eventually, even if they never say it that way.

after a while, the advice starts looping.
the insights plateau.
the recommendations stop being surprising.
the product stops feeling like it is learning you.

that’s when engagement tends to soften.

because real personalisation is much harder than most of the category admits.

it would mean:

→ understanding your personal context
→ learning your specific patterns over time
→ remembering what you’ve already tried
→ tracking what actually changed for you
→ and recalibrating as your life evolves

that is a very different product from a segmented health or wellness app with a slick onboarding flow.

it is much harder to build.

but it is also much more valuable building something that overcomes the personalisation illusion.

#heathtech #healthai #wearables #personalisation

i built a fully custom headless e-commerce frontend for my DTC supplement brand over the weekend. i'm not a developer. i'm not a designer. just a product guy with a Claude max subscription.

let me explain.

i launched Marine Organics in 2022 as an Amazon-first brand side hustle, selling organic sea moss capsules. for the past 3.5 years Shopify and DTC were always an afterthought - a default theme we never touched.

last year we went all in on a rebrand. new logo and packaging logo by Moritz Dengler-Kingerter. we commissioned South African based artist Joanita Cillié and photographer Adel Ferreira to create these incredible sculptural compositions using real ingredients - sea moss, bladderwrack, lion's mane, shells - to bring out the essence of what we do: blending 100% organic ingredients into functional supplements built on our hero ingredient, irish sea moss.

the creative work was done. and then... nothing. for over a year.

we couldn't find the time (or budget) to brief designers, wireframe the UX, find a developer, go through rounds of revisions, and actually relaunch the site with the rebrand. the gap between vision and execution felt massive.

this weekend I sat down with Claude and everything changed. literally.

and in 24 hours, I built:
→ a custom React frontend with Tailwind CSS
→ headless Shopify integration (Storefront API, GraphQL)
→ live cart with subscription support (Subscribe & Save with selling plans)
→ Judge.me reviews pulled in via a custom serverless API
→ product pages with dynamic galleries, ingredient breakdowns, and mulit-month supply purchase options
→ a custom bundle selector & cart upsell module
→ journal/blog section connected to Shopify's CMS
→ EO redirects preserving all old Google-indexed URLs
→ deployed on Vercel with existing custom domain and auto-deploy from GitHub

no designer. no developer. no agency. just me and Claude jamming.

i am genuinely mindblown. every day a little more.

this is like a dream and I really do not want to wake up.

i have never felt more empowered as a builder.

i've always had ideas - but without coding skills, so many of them died at the ideation stage. that era is over.

times have changed. time to build.

grab yourself some sea moss if you're curious. or just poke around and let me know where my AI-generated code created some bugs - I'm sure there are plenty 😄

i hate to admit it, but i’ve been falling for something that i increasingly see across the wellbeing industry.

i call it optimisation theatre.

it looks like this:
the supplement stack.
the sunrise lamp.
the blackout blinds.
the journals.
the trackers.
the carefully engineered evening routine.

it looks like discipline.
it feels like control.

and to be fair, sometimes parts of it do help.

but i think there’s a trap hidden in all of this.

the routine can start to feel like the result.
the protocol can start to feel like progress.
the visible effort can start to substitute for actual change.

i’ve definitely caught myself doing this.

spending time refining the setup.
adding one more input.
buying one more thing.
tracking one more variable.

and asking a lot less often:

is any of this actually making my life meaningfully better?

that’s the uncomfortable bit.

because a lot of what passes for optimisation in wellbeing is really just participation in the culture of optimisation.

it gives you rituals.
artefacts.
a sense of identity.
being part of the "enlightened".
something to point to and say: look, i’m taking this seriously.

but real progress is usually much quieter than that.

which is maybe why this pattern is so easy to fall for.

and our category often rewards the appearance of effort more than the evidence of change.

and i think that distinction matters.

because you can get very good at performing wellbeing without getting much better at living it.

time to change that.

Chat is the new WYSIWYG.

For years, we built ever more sophisticated page builders and content systems around the idea that people needed interfaces to make change.

I spent my first years in tech building content management systems at Cognifide where we (all kudos to Fahim Salim & Tommy Yau) used to talk about three layers of change:

1. platform - slow, developer-led
2. UX/UI - still slow, design + dev
3. content - faster, more accessible, end users focussed

The whole point was to decouple them, so the slow cog wouldn’t hold back the parts of the machine that could move faster.

That thinking shaped a lot of digital experience architecture. AEM, Shopify front ends, Webflow, WordPress, Squarespace - different tools, different maturity levels, same basic paradigm:

WYSIWYG (what you see is what you get)

Complex interfaces. Templates. Training. Permissions. Workflows.

Now I find myself doing something that still feels slightly surreal.

I sit there and tell Claude what I want changed on the site.
What I don’t like.
Which asset to use.
What integration to build in the background.
Which checks to run.
Test it on localhost.
Push to Git.
Commit.
Done.

The interface is no longer the builder.
The conversation is the builder.

It still feels a bit like a fever dream.

Please don’t wake me up.

.
.
.
#ai #claudecode #webdevelopment

retention in consumer health is notoriously difficult. but i increasingly think it is less a retention problem than a meaning problem. solve the meaning part, and you solve a lot more of the retention part.

working in health tech natually makes you talk to people about their wellness routines - the gadgets they use, the protocols they follow, the apps they’ve tried and abandoned.

the number one thing i hear isn’t
“i need more data.”
it isn’t
“i need a better device.”
it isn’t even
“the recommendations are wrong.”

it’s this:
“i don’t know what any of it means for me.”

and i think that gets to the heart of what is still missing in the consumer health stack.

we’ve built a strong hardware layer: wearables, cgms, blood tests.
we’ve built a strong software layer: apps, dashboards, scores, trends.

but between raw numbers and real understanding, there is still a void.

users end up with what i call *numbers without narrative* - data points that never quite connect to the story of their actual life.

so they do one of three things:
1. they obsess over the numbers,
2. they ignore them, or
3. they churn.

and i think the retention data reflects that.

Business of Apps puts health & fitness app retention at roughly 3% by day 30.
the subscription picture is more nuanced. in the RevenueCat charts i was looking at, median health & fitness monthly subscriber retention is still around 19.2% at 6 months and 8.5% at 1 year. annual plans retain around 25.4% at 1 year. weekly plans are far weaker.

to me, those numbers don’t contradict each other. they explain the opportunity.

people are clearly interested enough to download.
some are motivated enough to pay.
but too few get to a level of understanding that feels personally valuable enough to stay.

that’s why i think the next winner in consumer health won’t just have a better sensor or a prettier dashboard.

it’ll be the company that builds the best interpretation layer.

the one that helps people understand what their data actually means for them - early enough, and clearly enough, to change the experience from information to insight.

i truly can’t stop thinking about that gap. curious whether others working in health, ai, or consumer product are seeing the same thing.

one thing i’ve noticed from paying much closer attention to my own health over the last few years:

i keep wanting simple rules.
and my own body keeps refusing to give them to me.

the same workout that makes me feel great one week can leave me flat the next.
the same amount of sleep can feel totally different depending on what else is going on in life.
the same routine that seems to “work” for a month suddenly stops feeling right.
i think that’s part of why so much health advice ends up feeling unsatisfying.

we still want clean answers in a system that is much more variable than we’d like to admit.

because the reality is: the same input does not reliably create the same outcome in every person. and often it doesn’t even create the same outcome in the same person across different moments of life.

that’s why i’ve become more sceptical of health products built around averages, benchmarks, and generic definitions of “optimal”.

they can be directionally useful.

but they are not the same as understanding you.

the real question is rarely:
“what works for people?”

it’s:
“what tends to work for me, under what conditions, and how do i know?”

that is a very different job.

less about pushing recommendations.
more about helping someone build a better model of themselves.

what actually tends to improve my sleep?
what kind of stress actually throws me off?
what patterns are real for me, and which ones am i just projecting onto the data?

that feels like a much more honest definition of personal health.

not perfect optimisation.
better self-understanding.
internal awareness.

it’s one of those things i genuinely can’t stop thinking about - curious whether others feel the same.

#healthtech #wearables #ai

your watch tells you your hrv dropped overnight. what, exactly, are you supposed to do with that?

that question has been stuck in my head ever since writing yesterday’s post.

because this is where most health products still break: not at measurement, but at interpretation.

the device can tell you something changed.
it usually can’t tell you why it changed.
and it definitely can’t tell you what that change means in the context of your actual life.

it doesn’t know you had a difficult conversation late at night.
it doesn’t know the stress in your body is excitement about a new project, not overload.
it doesn’t know your fatigue is emotional, not physiological.
it doesn’t know that yesterday’s meeting mattered more than your resting heart rate.

so we end up with a strange product experience: incredibly sophisticated measurement, wrapped in surprisingly shallow meaning.

that, to me, is the current *interpretation gap* in health-tech.

and it has big downstream consequences.

when people don’t understand what a score actually means for them, it becomes one of three things:

a ritual,
a source of anxiety,
or something they eventually ignore.

which is why i increasingly think the next layer of value in this category won’t come from better sensors.

it will come from products that can connect the signal to the story.

not just “something changed.”
but “here’s what may be happening, here’s why it matters, and here’s what to pay attention to next.”

that feels much closer to understanding.

#healthtech #wearables #health #ai

i’ve worn a fitness tracker every day for the last three years.

that’s 1,095 days of sleep scores, hrv readings, step counts, recovery metrics, resting heart rate trends, and all the other little signals we love to obsess over.

don’t get me wrong, i love my ŌURA ring. but if you asked me what i’ve actually learned about my health from all that data, i’d struggle to give you a very good answer.

sure, i know my hrv trends higher on rest days.
i know my sleep score drops when i travel.
i know my resting heart rate goes up when i drink alcohol.

but that’s not really insight.
that’s observation.
and observation without interpretation is just data hoarding.

i think that’s the uncomfortable truth at the centre of a lot of health tech right now: many of us have become very diligent trackers, but not necessarily better learners.

we’ve somehow convinced ourselves that if we just collect enough signals, clarity will follow.

but it doesn’t.

clarity doesn’t come from more data.
it comes from better interpretation of the data you already have.
it comes from context.
from pattern recognition.
from knowing what matters, what doesn’t, and what to actually do next.

the more time i spend building in this space, the more i think the category has a fundamental design problem:

we’ve built some incredibly good tools for collection,
and still far too few tools for comprehension.

we are very good at helping people track.
we are still not very good at helping people understand.

and that feels like the real opportunity.

anyone else sitting on years of health data they’ve measured very carefully, but not actually used very well?

...and yeah, for all of you sleep score obsessed folks, don’t shame me - I'm trying.

#healthtech #wearables #digitalhealth

less hype, more signal //
after openai, anthropic, amazon, and now Perplexity, ai-powered health is starting to look less like a side experiment and more like an emerging product category.

it also makes one bigger shift feel increasingly obvious:

*we are moving from the age of information to the age of understanding.*

for years, digital health mostly gave people more access to data, more articles, more dashboards, more fragmented signals.

now the opportunity is different.

not just surfacing more information, but helping people interpret it.
not just showing the numbers, but explaining what matters.
not just tracking health, but making it easier to understand what to do next.

that, to me, is the more interesting implication of all these launches.

the winner here will not be the company that can answer the most health questions.
it will be the one that knows when to answer, how to ground the answer in real personal context, how to communicate uncertainty, and when to hand off to a real clinician.

health is one of the clearest tests of whether ai can move beyond information retrieval and into real-world understanding.

because the real product challenge is not:
“can we summarise your data?”
it is:
“can we turn fragmented, personal, high-stakes information into something genuinely useful without creating false confidence?”

that is a much harder problem.
and a much more important one.

sober AI takes ///
just because you can now build it, doesn’t mean you should.

ai coding tools are removing a real bottleneck. ideas that used to take weeks to test can now be shipped in hours.

that is genuinely exciting.

but it also creates a new risk: people start confusing the ability to build with the judgment to build well.

ai can help generate code. it does not replace human intuition, taste, or product sense.

and that matters because customer experience is still a very delicate thing. the best products are rarely the ones that do the most. they are usually the ones that remove friction, stay focused, and make the experience feel simple.

when shipping gets easier, the temptation is to add more.
more features. more flows. more ideas. more complexity.

but the core principle of good product experience probably doesn’t change:
keep it simple. make it easy. remove more than you add.

ai will absolutely unlock a huge amount of new products and experiments.

but simplification, restraint, and taste are still the real differentiators because building faster is not the same as building better.

the bottleneck is shifting from shipping to judgment.

the new risk is not that we can’t build enough. it’s that we build far too much.

sober ai takes ///
how do we fix users being blind to on-screen notifications?

people ignore pushes, badges, banners, and little red dots all day long. but far fewer people leave actual messages unread than ignore app notifications altogether.

that’s why one of the most tactical takeaways from the openclaw-nification of everything that is currently under way is not that every company should suddenly become a pure agent business.

it’s that messaging may be the better re-engagement surface.

not the whole product.
not the whole business model.
just the attention layer.

the implication is pretty practical:
use the channel people already check obsessively to trigger intent, then pull them back into the app for the deeper, higher-value experience.

messaging as the attention layer.
the app as the experience layer.

the interesting question is not whether every business becomes openclaw.

it’s which parts of the user journey should move into messaging because that’s where attention still exists.

that feels a lot more useful than repeating “everything is an agent now.”

something is changing (2/2)

something is changing, but let’s cut through the noise.

if building software becomes dramatically easier - and it is - we’re heading toward a supply–demand distortion.

the supply of software will explode.
the demand for distinct tools will not.
that’s when things get interesting.

when supply outpaces demand, capability stops being scarce. features stop being defensible. “we built this faster” stops mattering.

so what matters?

distribution. trust. possibly brand. definitely networks.

if anyone can build a solid product, the advantage shifts to those who already have attention and trust. the builder economy becomes less about building and more about who already has an audience that cares.

brand compounds again.
not brand as a logo. brand as accumulated trust. brand as default choice.
in a crowded landscape, people don’t evaluate every option. they reach for what feels known and safe.

and then there are network effects.
some parts of tech will feel commoditisation quickly. single-player productivity tools. internal workflows. personal utilities. those will multiply endlessly.
but networked systems are different.
you can vibe-code something that looks like instagram. you cannot vibe-code a billion users.

you can build a kanban board in a weekend. that doesn’t mean a fortune 100 company will rip out jira on monday. we’re not moving away from linear because someone built a clever agent-driven alternative on youtube.

this is where the hype-cycle loses perspective.

the recent software sell-offs make for good headlines. “saas is dead.” “the end of enterprise software.”

traditional software isn’t going anywhere overnight.

enterprises don’t pivot because a new tool is impressive. they optimise for reliability, compliance, integrations, and vendor stability. the slow grinding gears of large organisations aren’t easily accelerated - not even by powerful models.

change will happen.
just unevenly.

on the edges first. with individuals. with small teams. then gradually inside larger organisations.

so the real question isn’t “is software dead?”
it’s “where does abundance change the rules - and where do the old rules still apply?”

if supply truly outpaces demand, build with distribution in mind from day one.
build with brand in mind. build with network effects in mind. build where coordination between humans matters, not just computation.

and regulate your own response.

it’s easy to get swept up. it’s easy to dismiss it.

the truth is somewhere in between.
these are exciting times. the leverage is real.

but even ai, for all its power, cannot bend the most consistent force in business - the slow, compounding inertia of systems, incentives, and human behaviour.

those wheels turn slowly.

and that slowness might be the most underrated stabiliser in all of this.

something is changing (1/2).

i can feel it, and it’s making me question what i’m doing, how i’m working, and what progress even looks like right now.

the linkedin feed is loud again.
“revolution.”
“disruption.”
“the end of X as we know it.”

everyone is sharing the latest hype-cycle article. everyone is watching the “how i built this app in 24 hours” video. everyone is listening to the one exceptional operator on lenny’s podcast and nodding along.

and yet, almost no one actually does anything.

that’s not new.

the rarest commodity is - and always has been - action.

ai doesn’t magically change that. it might compress time. it might lower the barrier slightly. but it doesn’t turn spectators into builders. most people will still consume the narrative rather than participate in the work.

what has changed, very recently, is the quality of execution you can get from a single person.

tools like claude code and the new codex versions aren’t just “ai assistants”. they can reason across files, hold context, write and refactor real systems, and close the loop between intent and implementation. this isn’t about prompts anymore, it’s about delegation.

and that’s where things start to blur.

engineers are moving into product.
product people are moving into engineering.

people who used to orchestrate work are now shipping it directly.
everyone who is curious enough is becoming a builder.

from the outside, this makes the space feel impossibly crowded. every day there’s a new tool, a new app, a new “i built this in a weekend” post.
but is it actually crowded?

or does it just look crowded because the cost of shipping something - anything - has collapsed?
most of what we’re seeing is surface area, not depth. experiments, not commitment. demos, not endurance.

which brings me back to the uncomfortable part.

when execution becomes cheaper, faster, and more accessible, the excuses disappear. the bottleneck moves inward. clarity. taste. judgment. persistence.

ai isn’t creating a world where everyone suddenly builds. it’s creating a world where the gap between those who act and those who don’t becomes impossible to ignore.

and that’s the part that’s making me pause - reflect - and re-evaluate what i want to spend my time building next.

interesting how everyone’s watching the “ai heavyweights” sprint into creating the ultimate healthOS but little is said about the sleeping giants in this space.

OpenAI launched chatgpt health and just went bigger in healthcare (with the gates foundation). Anthropic is shipping healthcare-specific claude features and record connections. and this week Amazon pushed an “agentic” health assistant inside one medical.

but the real sleeping giants here might be Apple and Google. not because they have the loudest demos, but because they own the rails healthOS can run on.

apple’s moat is quietly unfair.
it starts with trust. people will share health data with the brand they already trust with privacy by default. apple has spent years making on-device processing and privacy a first-class product principle.
distribution is the second layer. iphone, watch, airpods, and os-level surface area like notifications, lock screen, widgets, and siri mean apple doesn’t need to acquire users. distribution ships in your pocket. sensors and continuity are another advantage. wearables aren’t a feature, they’re a constant data stream, and apple owns the most tightly integrated one. on top of that sits the health app and healthkit, acting as a permissioned system of record for personal health data.
apple can be late on models and still win by making the default health experience feel safe, simple, and sticky.

google’s moat is unfair in a different way.
android gives it global scale and play store distribution at a level nobody else can match. it also owns its own foundation model family with gemini and medical variants. on the infrastructure side, health connect is becoming the interoperability layer across health apps. and when people have a symptom or health question, search intent still defaults to google.

there’s a wildcard here. if apple deepens its partnership with google’s gemini for a future siri experience, the equation changes fast. apple’s product design, trust, and distribution sitting on top of google’s frontier models would be an extremely hard combination to compete with.

so what actually matters more in health: the best model or the best product people trust?

my bet is that trust and distribution beat the best model in healthcare. models will continue to commoditise. what doesn’t commoditise as easily is permissioned longitudinal data, being present at the right moment on a wrist, in an ear, or on the lock screen, and owning the workflow from reminders to logging to hand-offs with clinicians.

which makes the real question this: are apple and google snoozing, or are they hibernating until they decide it’s time to make their move in the biggest sector on earth?

As promised last week, here is the deep dive into my latest article on unmuted.ai: "Why Every Product Manager Building in AI Needs to Start Thinking About LLM User Analytics".

This one dives into a topic I believe every product manager working in conversational AI needs to start considering: How do we adapt our optimisation and data-driven approaches to this new, dialogue-driven conversational interface?

The way we’ve traditionally measured product success simply doesn’t translate to conversational AI. These interactions are dynamic, highly nuanced due to the underlying natural language input/output paradigm, and often require us to dig deep into the conversations themselves to understand user intent, satisfaction, and even frustration.

Here’s what I have covered as part of this article:

1️⃣ The Missing Signals in Conversational AI:
Traditional metrics like “clicks” are no longer enough. A lot of the insights about how users engage with your product are hidden in the conversational flow, embedded in sentiment, intent, and behaviour within the dialogue itself.

2️⃣ Adapting the AB(C)C Framework:
Acquisition, Behaviour, and Conversion metrics still apply but need updates. Behaviour now requires looking beyond surface-level engagement to uncover emotional and intent-driven patterns in conversations. In this new paradigm, conversations themselves become critical, hence adding a C for Conversation to the framework. Tracking what happens within user dialogues, such as key topics, unresolved queries, and repeated questions, offers a deeper understanding of how effectively the AI meets user needs.
Conversion? It’s no longer just about the button click, it’s often embedded in a user’s request or a sentiment expressed within the chat.

3️⃣ Unlocking Dual-Purpose Analytics:
LLM user analytics tools such as Nebuly aren’t just about optimisation. They also safeguard your users and brand by monitoring risks like hallucinations or inappropriate outputs while surfacing trends that help you build better experiences.

4️⃣ Getting Started with LLM User Analytics:
Practical steps for product managers include defining clear objectives (e.g., improving user satisfaction or strengthening guardrails), choosing the right tools to align with your goals, and establishing an iterative feedback loop to continuously optimise your conversational AI. These foundations set you up for both delivering value and safeguarding your users and brand.

The full article is live on unmuted.ai/resources.

If you’ve been thinking about these challenges or want to explore how LLM user analytics could transform your AI products, I’d love to hear from you.

Drop a comment, shoot me a DM, or just let me know your thoughts.

#LLMUserAnalytics #ConversationalAI #AIProductManagement #UnmutedAI

https://lnkd.in/eYY4wSBz

AI-driven products are revolutionising user experiences, and at the heart of this transformation lies conversational interfaces powered by LLMs. As a product manager, you’re tasked with steering these innovative tools to success. But here’s the catch: the traditional metrics and frameworks you’ve relied on for web and mobile analytics are not comprehensive enough for this new frontier.

If you’re building in AI and not yet thinking about LLM user analytics, you’re missing a vital piece of the puzzle. This emerging type of analytics is the key to unlocking user insights, improving your product, and ensuring its success in an increasingly competitive space. Much like web analytics defined the success of the early internet era, LLM user analytics is poised to do the same for conversational AI. Let’s dive into why it matters.

Why LLM User Analytics Matters for Product Managers

1. Driving Product Improvements Through Data

Good product management thrives on data-driven decision-making, and LLM user analytics provides an additional layer of insights that traditional tools do not cover. New tools go beyond tracking engagement or retention, offering a granular look at conversation quality, user sentiment, and intent. Imagine understanding not just how users interact with your AI but also why they’re engaging and where the experience falls short.

For example, is your AI correctly interpreting user intent? Are conversations leading to successful outcomes? These insights are often hidden within the nuances of the conversational flow and sentiment, requiring a different type of analysis. Extracting these insights allows you to refine flows, optimise prompts, and make iterative improvements that align with user needs.

2. Introducing the ABCC Metrics Framework

Traditionally, product managers relied on ABC metrics, Acquisition, Behaviour, and Conversion, to measure the success of their products. While Acquisition metrics (how users are acquired and where they come from) remain relevant in the conversational AI paradigm, Behaviour and Conversion require a significant shift.

• Behaviour (B): Simple click and engagement metrics are no longer enough. Much of the insight into how users engage with conversational AI is hidden within the conversational flow, including sentiment, tone, and intent.
• Conversation (C): In this new paradigm, conversations themselves become critical. Tracking what happens within user dialogues, such as key topics, unresolved queries, and repeated questions, offers a deeper understanding of how effectively the AI meets user needs.
• Conversion (C): Unlike traditional interfaces where conversion might be tracked via a CTA click, in conversational AI, conversions are often embedded within the conversation itself.

Adapting the ABC framework to ABCC highlights how LLM user analytics is essential for uncovering user intent, enabling real-time optimisation, and building more intuitive conversational flows.

Consider a future where an LLM-powered customer support assistant for an e-commerce brand doesn’t just answer product questions but actively assists users in completing purchases directly within the chat. To enable this, product managers need to analyse user dialogues using LLM user analytics tools to identify signals of purchase intent. They can then refine the AI to act on these signals, delivering a seamless, proactive user experience.

3. Balancing Risk and Value Creation

AI product managers operate in a unique space where managing risk is just as critical as delivering value. One misstep, an inaccurate response, a privacy breach, or an inappropriate output, can jeopardise trust and damage your brand. At the same time, overly restrictive guardrails can stifle the AI’s potential to delight and engage users.

LLM user analytics helps you navigate this balance by providing metrics that address both sides of the coin. For instance, monitoring accuracy and safety metrics protects users while also identifying areas where the AI could better meet their expectations.

4. Proving ROI to Stakeholders

As AI investments grow, so does the pressure to demonstrate their return on investment. Product managers need robust analytics to build a compelling ROI narrative. By showcasing how conversational AI improves user satisfaction, reduces friction, or boosts operational efficiency, you can build a case for continued investment. LLM user analytics equips you with the data to answer the tough questions from CEOs and CFOs.

Core Capabilities That Product Managers Need from LLM User Analytics

1. Conversation Quality Metrics

A great AI product starts with strong conversational quality. Metrics like turn completion rates, success rates, and sentiment scores highlight where users are dropping off or feeling frustrated, offering insight into bottlenecks. Advanced analytics such as intent recognition and topic modeling go a step further, identifying patterns in user needs and intent to refine interactions and deliver better outcomes.

2. Real-Time Feedback and Iteration

Iterative improvement is key. Regular analysis of conversation logs, flagged interactions, and sentiment trends ensures that product managers can make data-driven updates to improve flows, prompts, and overall user satisfaction. The goal is agility, not immediacy, steady refinements that align the AI with user expectations.

3. Balancing Guardrail Monitoring with Value Identification

LLM user analytics doesn’t just track what users are doing; it also monitors how the AI is behaving. Tools that enable topic analysis are particularly valuable here. They can identify sensitive topics that require additional review to ensure guardrails are working, while simultaneously uncovering areas where users are seeking information or functionality that the AI cannot yet address.

Practical Steps to Get Started with LLM User Analytics

1. Define Clear Objectives

Start by identifying specific goals for your LLM user analytics efforts. Are you aiming to reduce churn, improve conversational success rates, or strengthen your guardrails? Your objectives will guide the tools and processes you need.

2. Select the Right Tools

Choose tools that align with your needs. Nebuly, for example, excels in analysing user intent and sentiment, while other platforms may focus on guardrail metrics or real-time performance monitoring. Deciding between building in-house capabilities or leveraging these established solutions is key to scaling efficiently.

3. Establish a Feedback Loop

Analytics without action is just noise. Ensure you have processes in place to analyse data, identify opportunities for improvement, and implement changes. By iterating on insights, you create a feedback loop that continuously optimises your AI’s performance, keeping it aligned with both business objectives and user needs.

Final Thoughts: Navigating the New Era of Product Management

The age of conversational AI demands a new playbook for product managers. LLM user analytics is the compass that helps you navigate this evolving landscape, offering the insights needed to improve user satisfaction while safeguarding your brand.

If you’re exploring this space and need guidance, I’m here to help. Send an email to mail@unmuted.ai or find me on LinkedIn to discuss how LLM user analytics can become a core pillar of your product strategy.

Let’s keep the conversation going, unmute your AI and unlock its full potential.

❗ The fastest way to fail with any GenAI project is to bring your deterministic mindset when assessing the risk and value of a generative, or rather, probabilistic technology. In the world of GenAI, there are no “guarantees.” We need to shift our thinking towards a confidence and probability-based approach to risk management.

In traditional tech, we’re used to controlling input and expecting predictable, controlled outputs. But with generative AI like LLMs, we only influence the outcome through input parameters and careful training - yet, the results remain inherently unpredictable. This shift changes everything about how we monitor, manage, and measure risk.

📃 This post marks the next and final piece in my series covering insights from my long-form article, “Monitoring LLM Performance to Manage Model Risk and Measure User Satisfaction – Two Sides of the Same Coin,” which I shared two weeks ago.

So how do we navigate this new landscape without falling into either extreme: overly restrictive or too loose? This is where performance monitoring as a dual-purpose tool comes into play.

⚖️ Monitoring isn't just for catching mistakes; it’s also a window into opportunities for creating value. AI that’s over-regulated might be safe but dull, lacking the engaging qualities users expect. On the flip side, overly lenient approaches may create richer interactions but risk undermining trust or compliance. The sweet spot? Adaptive performance metrics that evolve as your AI matures, striking a balance between safety and value.

Key takeaways for building a balanced strategy:
1️⃣ Don’t dismiss early-stage risk management as overly cautious, it’s the groundwork for future trust and value.
2️⃣ Prove success by delivering value and safeguarding users and the brand.
3️⃣ Repurpose risk management tools for user insights to enhance experience.
4️⃣ Trust the process; once safety concerns are settled, optimisation becomes the next focus.

🔇 The biggest risk is not knowing how users engage. Unmute those interactions to find the balance that makes your AI both compliant and compelling.

Companies that get this balance right will lead the next era of AI-powered user engagement. If you’re navigating this balancing act, I’d love to hear your thoughts. The full article dives deeper and is available at unmuted.ai/resources (link in the comments).

Let’s keep this conversation going ⬇️ ! #LLMMonitoring #LLMEvaluation #Evals #GenAI #RiskManagement #AI

Last week, I shared that I’d be breaking down my latest article on LLM performance monitoring into a series of shorter LinkedIn posts.

Today, let’s dig into the first part:

⚖️ How balancing risk management and value creation is essential for successful AI deployment.

When working on AI projects, you quickly learn that managing risk isn’t just about compliance or ticking boxes - it’s the bedrock that allows us to build AI systems that users, and, if you operate in an enterprise environment, your stakeholders can trust.

But here’s where it gets interesting: the same metrics that keep your LLM safe, accurate, and compliant can also act as indicators of user satisfaction. It’s a dual-purpose approach that makes a lot of sense if you think about it.

So, what does ‘risk’ really mean in the context of conversational AI? It’s not just operational issues; it’s a broader picture that covers:

🧮 Accuracy: Making sure your AI doesn’t hallucinate or provide incomplete information that could damage trust. Ultimately, ensuring safety by avoiding inaccuracies that could lead to harm is the primary goal.

🔒 Privacy and Compliance: Ensuring user data is handled responsibly, aligning with local and international regulations.

📰 Reputation: One wrong answer can do more than just confuse a user, it can impact how your entire brand is perceived.

In the early stages of deploying AI, it’s normal for risk management to take centre stage. After all, we’re dealing with generative tech that’s not as predictable as traditional, deterministic systems. This initial focus sets a strong foundation, ensuring that your AI behaves as expected and aligns with your brand values and regulations. And that’s where the magic lies: once you have risk management under control, you’re in a better position to pivot towards optimising for user value.

Want to know more about how performance monitoring can serve both risk and value? The full article is now live on Unmuted.ai/resources. Stay tuned for more follow-ups in this series!

And of course, I’d love to hear your thoughts. Are you also juggling these two sides of the AI coin? Drop a comment ⬇️ or reach out 📩 I’m always up for a chat.

#AI #LLM #LLMUserAnalytics #AIMonitoring #LLMEvaluation #AIRiskManagement

In conversational AI, as in real life, there’s always a balancing act: creating value for users while managing the risks that come with deploying a technology that’s no longer deterministic. This inherent unpredictability makes it a lot harder to safeguard than traditional tech. Interestingly, the same metrics that keep your conversational AI safe and reliable can often double as indicators of user satisfaction.

The challenge? Striking a balance in a typically resource-constrained enterprise environment - building evaluation capabilities that not only manage the inherent risk but also ensure that users derive real value from your conversational AI product.

When I first started working in conversational AI within Unilever, I saw firsthand how prioritising risk management was essential to protect users and brand integrity, especially for a large publicly traded company where a single PR misstep could tank its share price.

Now, as I continue my journey of building conversational AI interfaces in a high-stakes field like health tech at Flo Health, I continue navigating this balancing act and I’d like to share a few observations on why businesses should avoid separating risk and value in their LLM evaluation efforts.

Let’s start by exploring why performance monitoring can address both these critical objectives.

1. Understanding Risk in Conversational AI: Accuracy, Privacy, and Reputation

The shift from deterministic to generative technologies is a fundamental one, altering how we evaluate and manage systems. Unlike traditional, rule-based systems that deliver predictable outcomes based on predefined inputs, generative models - like large language models (LLMs) - produce responses that can vary widely depending on context, user input, and underlying data. This inherent unpredictability means that generative AI requires a different evaluation approach.

In deterministic systems, evaluations focus on ensuring that inputs consistently produce expected results. But with generative AI, especially in conversational interfaces, evaluating performance is more complex and nuanced. Responses can be contextually accurate or may “hallucinate” - generate plausible but entirely incorrect information.

So, when we talk about “risks” in AI, we’re not just discussing operational quirks; we’re talking about potential failures that could impact trust, compliance, brand image, and even your company’s share price.

Here are three key risk areas for any team building AI-driven interfaces:

1. Accuracy Risk: At its core, accuracy risk concerns the potential for AI to either “hallucinate”, generate convincing but entirely wrong information (untrue) - or fail to provide essential information (incomplete). In cases where users rely on AI for critical insights, these errors can erode trust, damage the brand, and, in the worst case, even harm users.

2. Privacy and Compliance: Privacy compliance is essential in industries handling personal or sensitive information, such as healthcare. Just as with any data-driven product, AI systems must align with local data and AI policies. Regulatory breaches here can lead to hefty penalties and long-term reputational damage.

3. Reputational Risk: Perhaps the most visible risk of all. Imagine an AI that occasionally delivers responses that seem off or insensitive. Even minor missteps can lead to a wave of user mistrust and harm the brand’s reputation, especially in sensitive sectors.

2. LLM Evaluation from a Risk Management Perspective

In the early stages of deploying conversational AI, companies understandably place heavy emphasis on risk management. After all, products and experiences are built on a new type of technology that’s largely untested and bears higher risk due to its generative nature.

Risk management within conversational AI isn’t simply about tracking accuracy; it’s about proactively measuring performance across critical dimensions like safety, bias, compliance, privacy, and reputation - before you go live, and especially after you’re live and users actually engage with your product.

Establishing tools and processes to evaluate these risks forms the backbone of responsible, resilient AI deployment, creating a foundation where the AI aligns with brand values, complies with laws and regulations, and meets user expectations.

Here’s a breakdown of key risk categories and the types of monitoring required:

1. Safety Evaluations: At the core of AI risk management is keeping interactions safe by preventing harmful or inappropriate responses. Techniques like “LLM-as-a-judge,” where one model evaluates another’s responses based on predefined safety criteria, can be effective and scalable but may falter when handling nuanced or complex rule sets. If two human reviewers struggle to agree on rule compliance, the LLM-as-a-judge approach may also fall short.

While this method can scale well for high-volume applications, I recommend carefully considering whether you’ll be able to fully automate the evaluation with this approach or whether a different/hybrid approach might suit your case better.

2. Accuracy Testing: Accuracy is essential for maintaining trust. It’s not only about correct answers but also about identifying when the model might be unsure or prone to hallucination. Regular testing against curated datasets helps track accuracy rates, while spot-checks or human evaluations on sensitive queries can reveal where the model’s confidence might outstrip its actual knowledge.

Your architecture plays a key role here: LLMs using RAG architectures often show reduced inaccuracy due to greater control over response content. However, even RAG architectures are susceptible to hallucinations, especially when queries fall outside the database’s scope. LLMs actually behave very similar to humans: a well trained human with high integrity will admit when he or she doesn’t have the answer to a question, others will confidently try to make something up. I suppose we’ve all been there.

3. Privacy and Compliance Monitoring: Privacy is especially vital when handling personal information. Regular audits and automated compliance checks help ensure data isn’t stored or processed in ways that could breach regulations. For example, businesses can implement filters or specific prompts to prevent the AI from discussing or storing sensitive topics. Although privacy audits can be resource-intensive, they are essential for maintaining user trust and regulatory compliance.

Implementing a robust PII data-scraping capability is also highly recommended. Such a tool scans and detects PII data within user dialogues, anonymising it before it gets stored. This step is crucial because, even if your conversational AI is not intended as a data collection tool, users may still enter personal information into the dialogue. Consequently, this data could end up in your logs, potentially putting you in breach of legal regulations.

Algorithm-based tools like Presidio are available for this purpose, but in my experience, an LLM-based approach can work even better, as it can detect more nuanced instances of PII without relying on a predefined library, which is almost always out-of-date by default.

4. Reputational Risk Management: Sometimes, a model’s responses can stray into ambiguous territory, potentially damaging a brand’s reputation (think McDonald’s, Google, AirCanada). Here, the key is to establish the same crisis response mechanisms that your company likely already has in place for other parts of the business.

The process is simple: identify and acknowledge the issue, take responsibility, over-deliver on the fix, and recognise that there will always be some folks “trying to get you.” But when deciding between delivering value to your users and exposing yourself to the risk posed by these PR vultures, always opt for your users.

5. Balancing Scalable vs. Non-Scalable Evaluations: When managing risk in conversational AI, there’s a constant balance between scalability and precision. Popular methods, like using the LLM as a judge, are excellent for scalable, quick evaluations but aren’t foolproof, especially in sensitive contexts. In these cases, non-scalable steps like manual reviews or risk- and topic-based analysis for highly sensitive queries are invaluable.

The goal is not to choose one over the other but to use each in tandem, ensuring that high-risk interactions are covered by a mix of scalable tools and meticulous human oversight.

3. The Balancing Act: Why Performance Monitoring Must Serve Dual Purposes

Here’s where the balancing act comes into play. The best performance monitoring setups allow AI to stay flexible enough to create satisfying user experiences while remaining safe and reliable. In other words, monitoring isn’t just about pinpointing mistakes or risks; it’s about identifying opportunities to add value.

The beauty of this approach lies in its adaptability. As conversational AI matures, its monitoring systems should evolve, reflecting the shifting needs of both the company and users. If AI is too regulated, it might feel safe but bland, lacking the engaging qualities that make interactions memorable. On the other hand, giving the AI too much freedom could enrich conversations at the expense of safety. The key is to adapt incrementally, using performance metrics to guide these adjustments.

4. Key Takeaways: Building a Balanced Evaluation Strategy for Your Conversational AI

How can businesses balance risk and value in a way that benefits both brand and user? Here are a few practical takeaways:

1. Don’t fight the focus on risk management: The early stages may feel overly cautious, but proving that your AI can be trusted to manage risks will pay dividends in the long run. Risk management isn’t just a box to check; it’s the foundation on which all future value is built.

2. Proving success with AI means succeeding on two fronts: delivering value to users and keeping them safe. An AI that doesn’t protect its users or the brand will ultimately undermine its own value. This dual accountability is what builds sustainable trust with both users and stakeholders.

3. Use risk management tools to gather user insights: Tools initially intended for risk assessment can often reveal opportunities to enhance the user experience. For instance, topic analytics can identify sensitive areas for safeguarding, but they can also highlight trending topics where users are interested but where the AI currently lacks depth. These insights help you spot gaps in user experience that need attention.

4. Trust the process: Once initial risk concerns are under control, the focus will naturally shift toward optimising user engagement. This transition doesn’t happen overnight, but it’s inevitable. When the CEO or CFO eventually asks for the ROI on the AI, being ready with a well-monitored, data-backed approach will allow you to demonstrate tangible value, without being caught off-guard.

5. The biggest risk is not knowing what your users are doing: Whether it’s a website, an app, or conversational AI, understanding how users engage with your product is critical. Unmuting those interactions wink is the only way to manage risks effectively while also optimising for real value. This dual-purpose insight provides the information needed to fine-tune both safety measures and user engagement.

5. Final Thoughts: Unmuting the Full Potential of AI

Balancing risk and value isn’t just about covering bases; it’s about using those insights to create impactful AI experiences. The future of conversational AI lies in its ability to adapt, learn, and continuously improve, using real-time insights to enhance user satisfaction and operational safety.

For anyone serious about conversational AI, this isn’t a nice-to-have, it’s essential. Because in the end, companies that strike this balance will lead the next evolution of AI-driven engagement.

With Unmuted.ai, I’m committed to helping businesses find this balance, “unmuting” valuable conversational AI insights. When performance monitoring looks through both risk and value lenses, we unlock a path toward AI experiences that aren’t just compliant but compelling.

Agentic use cases are definitely nothing new but it still feels a little bit like black magic seeing LLMs "autonomously" complete tasks on someone's machine.

That feeling is no different when it comes to Anthropic's latest update introduces Claude 3.5 Sonnet and its new experimental computer use feature, which enables the model to interact with computers much like a human, navigating interfaces, clicking buttons, and typing text. This capability, though still in its early stages, has the potential to enhance workflow automation, as it allows the LLM to take immediate action based on real-time inputs. Supposedly Asana, DoorDash, and Replit are already exploring these possibilities by leveraging Claude for complex multi-step tasks​.

💭 Food for thought: In the article I shared yesterday, I mentioned how the current state of sentiment analysis in conversational AI is often too static, with insights typically acted on post-hoc. Claude's new agentic abilities, and any advanced agentic capabilities, point towards a future where the LLM might be albe to take action on sentiment in real-time, adjusting conversations on the fly. I guess we'll find out soon enough considering the pace at how thing develop!

Full announcement here > https://lnkd.in/gpMnWJEn

Latest article here > https://lnkd.in/g2EFTgiZ

In the world of web analytics, sentiment analysis has always been a useful tool for understanding how users feel about your brand - whether it's positive, negative, or somewhere in between. But here’s the thing: sentiment in the context of data derived from reviews or social media posts is somewhat static and mostly analysed async.

With the rise of conversational AI, we’re entering an entirely new phase of sentiment analysis - one that requires a deeper, real-time understanding of user emotions as they interact with AI-driven systems.

This evolution is not just about tracking how users feel about your brand at a high level; it’s about understanding their moment-by-moment intent, emotional state and level of satisfaction/frustration as they engage in a conversation with AI. If you decide to provide users with a conversational AI interface, it’s time to unmute (see what I did there) these interactions and capture the full spectrum of user sentiment.

1. Traditional Sentiment Analysis vs. Conversation-Based Sentiment in LLMs

Sentiment analysis of user interactions on the web and social media has traditionally been pretty straightforward - usually focusing on positive, negative, or neutral sentiments derived from static text, such as written reviews or social media posts. While this method has served its purpose, it falls short in today’s era of conversational AI.

Unlike static data, conversational AI unfolds dynamically. User sentiment can shift throughout the interaction, meaning businesses need a more sophisticated approach. In conversational interfaces, sentiment isn’t just an indirect measure of brand perception - it’s a direct indicator of the quality of interaction between AI and the user. Conversations evolve, emotions change, and subtle cues like tone and pacing become crucial for assessing the success of these interactions.

This shift calls for new methods of monitoring conversations and tapping into the rich insights that come from real-time, in-depth conversations between users and AI.

2. Challenges: Detecting Nuances Like Sarcasm and Complex Emotional States

Now, here’s where it gets tricky: conversational AI brings a host of new challenges to the table. One of the biggest hurdles is detecting the emotional nuances that arise during real-time interactions. Think about it - users don’t always express themselves plainly. They might throw in sarcasm, humour, or subtle frustration, which are harder for traditional models to pick up.

There are two major challenges here. First, emotions like sarcasm or confusion can easily slip through the cracks of basic sentiment analysis tools. Second, because users have much higher expectations when interacting with AI, they’re more likely to express both positive and negative emotions directly during the interaction. This means businesses need to be equipped to gauge real-time sentiment to ensure their AI is meeting user expectations.

To solve this, LLM-based sentiment analysis relies on more advanced linguistic models that can recognize and interpret these emotional complexities, ensuring that businesses can provide better responses based on real emotional understanding.

3. How LLMs Enable a Deeper Understanding of User Sentiment (Beyond Positive/Negative)

With advanced linguistic models, LLMs are moving sentiment analysis beyond the basic “positive” or “negative” categories. Tools like Context AI and Nebuly have started providing LLM sentiment capabilities allowing to capture signals such as intent, sentiment and a broader emotional spectrum.

While these tools are a good start, there’s still work to be done. Right now, they provide some capabilities, but the real challenge lies in acting on these insights in real-time. Fully detecting the subtle signals hidden in a user’s query or response, and then feeding that information back to the LLM to adjust the immediate response and/or the conversation flow on the spot, remains an area where more innovation is needed.

Ultimately, tools like these will need to evolve further to fully unmute (3rd time, I’m on a roll) the potential of conversational AI. The ability to detect rising frustration, for example, and immediately adjust the AI’s response or escalate the issue to a human agent, will be key to creating more personalised, responsive interactions.

As sentiment tracking dashboards become essential for businesses aiming to improve user satisfaction, the future lies in fine-tuning these insights to truly optimise the conversational experience in real-time.

4. Case Study: How Evolving Sentiment Analysis Improves Customer Experiences

Let’s look at an example from the e-commerce world. Imagine a customer service chatbot interacting with a user who’s trying to make a purchase but gets stuck. Traditionally, the business wouldn’t know if the user was frustrated until they had analysed the transcript post-hoc, or until the customer had left a negative review.

But with coupling an LLM-driven conversational customer support chatbot with advanced sentiment analysis, businesses can detect frustration as it happens - say, through repeated queries or escalated tone - and adjust the conversational flow or redirect the user to a human agent before the experience goes downhill. That can make the difference between a purchase and a churned customer. At scale, it can make the difference between a successful and a struggling brand.

5. Conclusion: Unmuting Your AI for the New Age of Sentiment Analysis

The world of sentiment analysis is evolving. It’s no longer just about tracking static emotions - it’s about capturing the real-time, nuanced emotional journey of your users as they engage with conversational AI. By leveraging the power of LLMs and advanced sentiment analysis tools, businesses will be able to truly unmute their AI, unlocking deeper insights that go far beyond the traditional sentiment metrics of the past.

At the end of the day, understanding user sentiment is key, but what truly sets businesses apart is how they act on those insights. As we’ve learned from web analytics, sentiment is just the beginning - what matters is how you use that data to continuously optimise your AI’s performance, creating personalised, real-time experiences that keep users engaged.

🔉 Earlier this week, I announced the launch of Unmuted Ai -a space where I will be posting and sharing my thoughts and experience from the emerging world of #LLMUserAnalytics over the next months , and now I want to deep-dive into the first article I wrote on:

🔮 What the History of Web Analytics Can Help Us Predict About LLM User Analytics.

The first thing to recognise? We’ve been here before. The rise of web analytics in the late 90ies and early 2000s was a game-changer for businesses trying to understand their digital presence. Before that? Websites were basically black boxes, brands knew users were landing on their sites, but had no idea what they were doing once they got there.

Sound familiar? It’s exactly where where many businesses are today with their new LLM-driven conversational interfaces.

In my first article, I explore the parallels between the early days of web analytics and where we are now with LLM user analytics. Learning from the past gives us a major advantage. Let's use it!

Here are a few key takeaways from the article:

1️⃣ Web analytics started with basic tracking but quickly evolved to uncover user intent and behavior. LLM analytics will follow a similar path as tools become more advanced.

2️⃣ Early adopters of web analytics gained a competitive edge, brands today can do the same with LLM analytics by investing early in tools that provide real insights.

3️⃣ Your web analytics tools are great at what they do, not for LLM user analytics though. Luckily companies like Context.ai and Nebuly are leading the charge in helping businesses unlock the true potential of their conversational AI.

4️⃣ Future trends: As with the evolution of web analytics, using LLM user analytics to generate insights is just the beginning, the true value comes from using those insights to fuel continuous optimisation and real-time personalisation through ongoing experimentation.

This article provides a short primer for anyone interested in understanding and capitalising on the rise of LLM user analytics. You can find the full article in the comments, and I’ll be sharing more insights in the coming weeks. ⬇️

Jump into the conversation, I’d love to hear your thoughts! 💬

#LLMAnalytics #LLMUserAnalytics #ConversationalAI #UserInsights #AI #WebAnalytics #UnmutedAI

I. Introduction: The Evolution of Web Analytics and the Rise of LLM Analytics

Back in 2015 I was involved in one of Europe’s largest Adobe Analytics implementation projects. By that time web analytics had already been around for more than 20 years. The first web analytics tools, such as WebTrends, had been introduced in the mid-1990s, followed by Omniture and Urchin, the precursors to Adobe and Google Analytics.

Despite decades of progress, many businesses still hadn’t fully embraced the importance of investing in tools to truly understand their websites, one of their most important sales and marketing channels.

Now, we’re at a similar crossroads with conversational interfaces powered by large language models (LLMs). These interfaces are creating new opportunities for businesses to engage with users, but they also present a challenge: just like early websites, they’re a “black box” when it comes to understanding user interactions. To thrive in this new landscape, we can learn from the past, applying the lessons of web analytics to the rapidly evolving world of LLM user analytics.

II. The Early Days of Web Analytics: Learning to See the Data

In the early days of the World Wide Web, websites were a mystery. Businesses knew users were visiting, but they had little understanding of what those visitors were actually doing. The first web analytics tools, such as hit counters and basic traffic trackers, focused on simple metrics like page views and site visits.

These tools gave businesses a glimpse into their audience’s behaviour, but they didn’t reveal much about why users were visiting, what they were looking for, or how effective the site was at meeting their needs.

As websites became more sophisticated, businesses needed deeper insights into user behaviour. Advanced analytics tools like Adobe Analytics and Google Analytics began to track user journeys, conversions, and engagement metrics. These platforms moved beyond surface-level metrics to help businesses understand the “why” behind user actions, allowing for optimization of content and user experience.

III. The Black Box of Conversational Interfaces Today

Today, we find ourselves in a similar situation with conversational AI. Chatbots, conversational interfaces and AI agents powered by LLMs are becoming integral to how businesses engage with customers, but the interactions they generate remain largely opaque. We can see that users are engaging with these systems, but understanding the intricacies of these conversations, what users are asking, how they feel, and whether their needs are being met, continues to be a challenge.

This is where LLM user analytics comes in. Just as web analytics unlocked the mysteries of early websites, LLM analytics will help businesses decode the conversations taking place in their AI-driven interfaces. By analysing the content, flow, and outcomes of these interactions and measuring implicit and explicit user feedback, businesses can optimise their conversational AI systems to deliver better experiences and drive greater value.

IV. Parallels Between Web and LLM Analytics: Lessons We Can Apply

One of the key lessons from web analytics is the need to move beyond basic metrics. In the early days, page views and hit counters provided a starting point, but they didn’t tell the full story. Similarly, simple metrics like interaction count or session length aren’t enough to understand conversational interfaces. Just as web analytics evolved to track user journeys, LLM analytics needs to map conversational paths, analyse user sentiment and understand intent at a deeper level.

Another important lesson is the role of tools and platforms in driving analytics maturity. The introduction of sophisticated web analytics platforms like Adobe and Google Analytics transformed how businesses approached online optimisation. Today, new platforms like Context AI and Nebuly are emerging to provide similar capabilities for LLMs, allowing businesses to track, measure, and improve their conversational AI systems.

V. The Importance of Investing Early: How Early Adopters Benefit

History shows us that businesses that invested early in web analytics gained a significant advantage. By understanding user behaviour and optimising their sites accordingly, these early adopters were able to attract more visitors, increase conversions, and outpace their competitors. The same will hold true for LLM user analytics. Companies that embrace these tools now will be able to deliver better conversational experiences, improve customer satisfaction faster and drive tangible business results.

VI. Future Trends in LLM User Analytics: Predictions Based on Web Analytics

Looking ahead, we can expect LLM user analytics to follow a similar trajectory to web analytics in several key ways. As conversational AI continues to evolve, personalisation will become increasingly important, allowing businesses to use real-time insights to tailor interactions to individual users. Just as web analytics empowered companies to understand user journeys and behaviours on their websites, LLM user analytics will allow businesses to optimise conversations for better user experiences.

A significant shift will occur once we move from descriptive analytics, which simply explains what is happening in conversational experiences, toward interpretive, predictive, and prescriptive analytics. In the early days of web analytics, businesses relied on basic metrics to understand their website traffic. Over time, the focus shifted to more advanced insights, such as predicting user behaviours, identifying what would likely happen next, and prescribing specific actions to improve outcomes.

The same evolution is now beginning with conversational interfaces, where we will move from analysing conversations retrospectively to using AI-powered tools that help us predict user needs and prescribe actions to optimise interactions in (near) real time.

However, insights, whether from web analytics or conversational interfaces, only provide value if they are followed by action. Generating insights alone incur costs for a business without delivering any measurable benefit. In the world of web analytics it took some time but eventually businesses learned that experimentation was key to realising the value of these insights.

The rise of A/B testing and other methodologies enabled businesses to test hypotheses, compare outcomes, and continuously optimise their websites and apps for better results.

The same principle is likely to apply to LLM user analytics. As businesses gather more insights into their conversational AI interactions, the value will come from using those insights that drive continuous optimisation and inform the next best action. Without action, there is no value, and without value, there is no benefit to the business.

Continuous experimentation, powered by insights, will be essential to driving improvements in conversational experiences, ultimately delivering greater satisfaction for users and measurable benefits for the business.

VII. Conclusion: From Web Analytics to LLM Analytics, The Journey Continues

The evolution of web analytics provides a valuable playbook for navigating the rise of LLM user analytics. Just as businesses once unlocked the power of user behaviour on websites and apps, they now face the opportunity to do the same with conversational interfaces. However, the key to success lies not just in gathering insights, but in taking action.

The real value of analytics has always been in its ability to inform decisions and drive continuous optimisation. For businesses ready to embrace this next frontier, the path forward is clear: analyse, experiment, optimise, and let insights guide your next best action for your conversational AI. Those who act on these lessons will be well-positioned to thrive in the future of AI-driven interactions.

I think there are not enough people talking about this or maybe I'm in the wrong room 🤷‍♂️

According to CB Insights, Midjourney is tracking $200M* in revenue. Generally not a newsworthy headline, but...

1. Midjourney has never raised any outside money.
2. Midjourney is run by only 8 engineers plus the founder David Holz and two support staff.
3. Midjourney doesn't try to make the news every day or runs extensive marketing campaigns but truely relies on product-led growth.

...and all this despite still offering a somewhat clunky and questionable UI through the Discord interface.

This is a testiment that building truly incredible products DOES NOT require vast amounts of outside venture capital contrary to the often common belief these days. What it really requires is a bunch of highly capable individuals focussed on building a best-in-class product that is so good that it sells itself.

At an valuation of $10B that would be a mindblowing valuation per employee of almost $1B 👀 let that sink in for moment. Just to put this into perspective, OpenAI is currently at a $58.2M valuation per employee.

This all sounds almost too good to be true. Is there a catch? They should be insanely profitable considering their lean team and product-led growth but what about the compute cost?

Any insights, let me know below. ⤵

*This number is from October last year so this is likely ever higher now.

Some call it the ChatGPT moment of AI music creation. 🎶

Whether you're a seasoned musician or simply a music lover, you should definitely check out Udio. This innovative platform enables users to create professional-quality music tracks from simple text prompts. Just specify topics, genres, and other descriptors, and Udio transforms them into your own custom-made tracks. 🎼

TL;DR: Write a prompt, be a little patient, and prepare to be amazed by your custom track.

I'm surprised by the quality of the output. As someone who's not musically inclined (whatsoever), my ability to craft a sophisticated prompt is minimal, the results from my first, simple input ("chilled melodic indie style version of Drunken Sailors") were impressive. The track featured thoughtful lyrics ("Would you sail with me with winds that aren't ours" 😮), a nice melodic acoustic guitar, and a good pace. Plus, you can remix tracks, tweak lyrics, add intros, and more. 🎚

Imagine what professional musicians and composers, or anyone with a bit of musical talent (definitely not me!), could achieve with such a tool.

Look, many people will jump to doomsday scenario outcomes for the music industry, claiming that these AI tools mark the beginning of the end for musicians, particularly those in the long tail, yet it's worth considering the immense opportunities Udio presents. These tools could usher in a new era of music creation, enabling artists to experiment with bold, unconventional ideas swiftly and creating completely new, cross-over music genres.

As many have said before me: AI won't replace people, people who know AI will replace people. 👨‍🚀➕🤖

Curious and beyond excited to see what people will do with it and what's to come for the music industry in the age of AI.

Check it out yourself - link in the comments - and let me know what you think.

#AI #Innovation #Udio #TextToMusic

I don't usually share work updates, but I'm truly excited about a recent project I've been involved with at Unilever. It's something special that I think deserves to be shared. 🌟

Two months post-launch, the Dove scalp + hair AI Therapist in the US, a consumer-facing, LLM-driven, fully compliant AI assistant, is reshaping how we connect with consumers using AI, demonstrating Unilever's leap into AI-driven consumer engagement. 🚀

We built a product that harnesses the power of OpenAI's GPT 3.5, embedded within a Retrieval-Augmented Generation (RAG) architecture, and integrates with our in-house AI Beauty Advisor (AIBA) diagnostics services. This project marks a move away from static, one-dimensional web interactions towards a dynamic, conversational-AI-driven experience. It paves the way for a future where web experiences are not just responsive but truly interactive and personalized, answering basic inquiries and generating micro web experiences tailored to more complex questions through simple chat interactions. 🤖

We're already witnessing promising early successes, with great engagement and completion rates in both chat functionalities and diagnostics. This is concrete evidence that AI can foster deeper, more meaningful interactions. 📈

The journey of bringing theoretical AI discussions into a tangible, live product for one of the world's largest brands has been nothing short of exhilarating. Venturing into a new, emerging domain has presented us with numerous challenges and setbacks, resulting in a steep yet immensely rewarding learning curve. 📘➡️🚀

A shoutout to the project sponsors Rachel Francois St. Fleur, Zoe Eungblut and Richard Jerrett for their relentless dedication and expertise bringing this to life. Special thanks to my manager, mentor, and friend, Fahim Salim, for your unwavering trust, guidance, and for giving me the opportunity to lead this journey from the start. Also thanks to Khushbu Varshney and the entire Sapient team for all the hard work from initial concept all the way to the final build. Working together to turn complex AI concepts into a live, consumer-facing tool has been an inspiring experience. 🙌

This project marks a significant milestone in how we envision the future of consumer engagement at Unilever, interactive, insightful, and, above all, personalized.

Here's to continuing our journey of innovation and learning in the AI landscape!

If you want to give the assistant a try yourself, link is in the comments.

#AI #LLM #Innovation #DigitalTransformation

Honestly, which names came to mind when you thought about big tech, innovation, and "Companies I Like To Follow" (let's call them CILFs, shall we?) in the past 20 years? I bet Microsoft hardly made anyone's list. So how have they turned from a geeky, clunky PC giant, to a cloud leader, to the #1 AI mover and shaker?

For many years in the early 2000s, Microsoft wasn’t really on anyone's radar when it came to big tech dominance and innovation. A failed smartphone and mobile OS (Hello, Apple !), no footprint in the social web whatsoever (What's up, Meta ?), and irrelevant shares of the search and browser market (Nice to meet you, Google .), Microsoft managed to reinvent itself through a focus on enterprise cloud storage and solutions, and now through an unprecedented focus and sharp execution in the AI space.

Satya Nadella has executed one of the most impressive big corporation course corrections of the past decades, and the stock price speaks for itself. When Satya took over as CEO in 2014, the stock was trading around the $40 mark, the same as 14 years earlier during the Dotcom boom. As I'm writing this, the stock trades at an all-time high of $415 per share. I'm neither a tech journalist nor a stock analyst, but that's pretty impressive if you ask me. People love a good old 10x story, don’t they?

So, what's next for Microsoft ? Well, everyone is probably aware of one of the smartest investment decisions (may I say "ever"?) into OpenAI last year. But did you know that Microsoft also led a $16m round investing in open-source darling, and arguably Europe's last hope to play any relevance in the future AI wars, Mistral ?

The deal includes bringing Mistral’s LLMs onto Microsoft’s Azure cloud computing platform. You know the platform that, according to Microsoft, 95% of Fortune 500 companies operate their business on. I anticipate this deal will also grant preferential access to Azure's computing resources. Both are obvious wins for Mistral , but why might this also be a smart move on behalf of Microsoft ?

Here’s 5 reasons why this is a smart move by Microsoft:

Strategic Diversification: Investing in Mistral AI allows Microsoft to balance its portfolio between open-source and proprietary technologies, offering flexibility to adapt to market and regulatory changes.
Innovation Access: By aligning with Mistral AI, Microsoft taps into the innovative potential of Europe's AI scene, ensuring a fresh influx of ideas and technologies to stay ahead in the AI race.
Expanded Market Reach: Bringing Mistral’s LLMs to Azure widens the service appeal, attracting a diverse clientele looking for open-source solutions, and cementing Azure’s position as a versatile cloud platform.
Regulatory Foresight: Anticipating future regulatory landscapes, especially in the EU, this move positions Microsoft to navigate potential restrictions on closed-source AI models, ensuring they remain a step ahead in compliance and continued access to the European market.
Community Engagement: Supporting open-source projects like Mistral AI enhances Microsoft's reputation within the developer community, fostering goodwill and encouraging collaboration on its platforms.

Also, $16m is barely pocket change for Microsoft these days and is not even noteworthy when looking at other recent AI funding rounds.

It’s an exciting time to be working in tech, I think. It feels like someone hasoxygenated a stuffy old room in the last 12 months as we observe established sectors being reinvigorated with new ideas and partnerships.

#AI #Microsoft #OpenAI #MistralAI

Interesting glimpse into the future of AI Interaction Design 📲. The Browser Company really trying to innovate with Arc in a space that has been quite boring for the past two decades.

With a simple pinch motion the Arc browser summarises the content of any webpage in a slick UI, overlaying the original page.

Will be interesting to see how these AI-driven interactions will impact the business models of the underlying content producers.

📉 A pessimistic view could be: Less time spent with on site > less impressions > less ad revenue > less money to spend on producing good content > quality of content descreases over time.

📈 An optimistic view: Content creators will find new ways to monetise their content. Maybe less through the more obvious ways such as subscription models or more content behind paywalls but more licensing focussed models following the likes of Reddit, Inc.

#AI #Innovation #ContentCreation #FutureOfWeb

I’m honestly often overwhelmed with the amount of information we can consume nowadays and it’s often challenging to know what deserves our attention. However, when seasoned Silicon Valley executives leave their high-profile roles at big tech firms to get back into start-up land, it’s a signal I think we can’t afford to overlook.

So when former Salesforce Co-CEO Bret Taylor and former Google Labs VP Clay Bavor join forces to launch Sierra, it’s wise to have a closer look.

Sierra's focus on conversational AI as a new form of interacting with the internet suggests a hopefully long-awaited evolution from the static, impersonal web experiences of the last 30 years. The duo believes that conversational AI will become the main way people interact with brands, offering a new, efficient form of communication for various customer service aspects. The technology allows customers to input queries in a search box, with AI understanding and acting upon these requests by connecting to transactional back-end systems. Despite challenges like integrating with legacy systems and managing data privacy, Sierra is focused on mitigating risks and even exploring moving away from a standard subscription-based pricing model that is common for most SaaS companies nowadays, to a more outcome-based pricing models. Curious to see if that sticks. Definitely an interesting approach to challenge the (huge and wealthy) incumbents in this space who will certainly not sleep on this and launch their own solutions.

Seeing innovation, capital and new companies moving into this space makes me even more excited about a quite transformative project I've recently been part of at Unilever. I’ll be sharing more about it soon.

For more details on Sierra, check out the full article on TechCrunch: [How Sierra is rethinking customer experience in the age of AI](https://lnkd.in/e8bYz5Hr).

#AI #AIAgents #Innovation #CustomerExperience #FutureOfTheWeb

Another day, another model, another milestone!

Google launches new open source LLM model Gemma, which outperforms Mistral AI 7B AND is available for commercial use according to Philipp Schmid from Hugging Face.

Who can keep up with this anymore? 😅

My initial thoughts on the factors that might have influenced Google's decision to follow fellow big tech buddy Meta's release of Llama 2 last year and release Gemma as an open-source model:

1. Community Collaboration: Open sourcing allows for broader community collaboration, harnessing the collective expertise of developers, researchers, and enthusiasts worldwide to improve and innovate on the models.

2. Transparency and Trust: Releasing models open-source can enhance transparency, allowing for more rigorous ethical scrutiny, bias mitigation, and trust-building among users and stakeholders.

3. Innovation and Competition: By making advanced models like Gemma openly available, Google contributes to a competitive landscape that encourages innovation, potentially leading to the development of even more advanced AI technologies.

4. Strategic Positioning: As AI governance and ethical standards become increasingly important, contributing high-quality open-source models positions Google as a leader committed to responsible AI development.

Read more >>> https://lnkd.in/ebjMqEWU

#AI #Google #geminiai #gemma

https://lnkd.in/eWb4NYN7

Ali Abdaal continues to impress with his openness and honesty about his journey as both a creator and a CEO and this latest video offers another valuable insight into the realities of this space.

One aspect that strikes me is the continued heavy reliance of creators on brand deals, which often necessitates a relentless output of content to meet payroll for their staff and maintain a profitable business. This constant cycle can sometimes overshadow the passion that often initially drives creators.

I firmly believe that creators should explore building more sustainable assets to support their operations. Such assets can provide a financial buffer, reducing the pressure to accept every brand deal that comes their way. This shift from feeling "I have to" create content for financial reasons, back to "I want to" create for passion, is crucial.

As Ali aptly put it, running your creator business with strategic constraints is key. It's about balancing the drive for profit with the original passion for creating. Thanks again for sharing your journey – it's truly enlightening.

#creatoreconomy #creators #creatorcommunity #creatorpreneurs

Who would have thought two months ago that there is even the slightest chance for a decrease of mobile media consumption?

Another proof point that (1) situational context rules, (2) the only thing to know for sure is that everything is always shifting and (3) crises usually just tend to accelerate already existing trends (see social gaming, remote working & collaboration, streaming, etc.)...it’s neither threat nor opportunity, it’s simply the reality of constant evolution.

#digitalstrategy #mediaconsumption #digitaltrends #coronavirus