William Kearns

Building AI for wellness @Purpose_AI + @uw, prev @Headspace

[William Kearns @kearnsw]

Most AI mental health apps treat chat as the product.

It’s not. Chat is the delivery mechanism. The product is behavior change.

If someone talks to your AI for an hour and feels heard but does nothing different tomorrow, that’s not coaching. That’s a confessional.

[William Kearns @kearnsw]

the biggest edge for builders right now isn’t keeping up with every new tool. it’s fluency.
and no, corporate doesn’t count. “circle back,” “zoom out,” and “take this offline” won’t cut it anymore.

https://t.co/UUz4YCVJRm

[William Kearns @kearnsw]

@LitRPGforum @chongdashu an ode to nethack

[William Kearns @kearnsw]

@Hesamation for those who’d rather grow their brain instead of rot it while waiting

[William Kearns @kearnsw]

built a spaced repetition learning CLI to fill the gaps between claude code requests

heavy AI use can atrophy your ability to recall and articulate ideas on your own

a few dozen cards per 2-5 min wait instead of doomscrolling or task switching helps keep me sharp for conversations with colleagues and AI

https://t.co/xM4k3QbFfR

[William Kearns @kearnsw]

built a spaced repetition learning CLI to fill the gaps between claude code requests

heavy AI use can atrophy your ability to recall and articulate ideas on your own

a few dozen cards per 2-5 min wait instead of doomscrolling or task switching helps keep me sharp for conversations with colleagues and AI

https://t.co/xM4k3QbFfR

[William Kearns @kearnsw]

libraries will continue to exist because reinventing the wheel is expensive

that doesn’t change with machines doing the labor since compute isn’t free

if anything the incentive to share increases since profit is limited when code can be generated cheaply

better to share working solutions than burn cycles on a million of the same apps, so we can reallocate that compute and human attention toward harder or more interesting problems

[William Kearns @kearnsw]

@samswoora always has been

[William Kearns @kearnsw]

@SahilBloom the said principle (specific adaptation to imposed demand) extends beyond the physical

[William Kearns @kearnsw]

@t_blom great for negotiating with insurance as well, do wonder how long they’ll accept inbound ai calls

[William Kearns @kearnsw]

@paraschopra a lot of non-engineers would disagree

[William Kearns @kearnsw]

@phuctm97 gym’s closing the days around christmas really messes with our habits. good time to remember frequency matters more than intensity for maintenance. thanks for the reminder to squeeze one in to get the ball rolling again before the new year rush.

[William Kearns @kearnsw]

@adonis_singh from my exp wouldn’t doubt opus to be the winner, but these type of single run tests are pointless. you’d need to run a sample from each model to understand the true mean performance / variance.

[William Kearns @kearnsw]

@burkov @forgebitz think the post is more about feature dev than full apps from a single prompt. one of the biggest unlocks i’ve experienced is to stop as much back and forth, by time you get to that point the context is polluted with wrong code / thinking tokens.

[William Kearns @kearnsw]

@forgebitz depends if you’re working at or beyond the frontier of the model, crud / tensor manipulations are enough for most apps. for more complex tasks it’s often better to go back and attempt to one-shot with additional context or multiple agents + git worktrees than to go back and forth

[William Kearns @kearnsw]

@Shpigford notion

[William Kearns @kearnsw]

@juliarturc Think the difference is Opus 4.5, before that most agentic output felt like slop. Think in general there’s a “boy who cried wolf phenomenon” due to non-stop hype with each model release for years. To many, this is a wake up that maybe this time is different.

[William Kearns @kearnsw]

We’re building this layer at Purpose to support our users in their daily lives. Great solutions don’t really exist yet.
Vector databases are solid infrastructure, but off-the-shelf products don’t have native methods for decay, inference, or distinguishing venting from belief.

[William Kearns @kearnsw]

Memory in AI is harder than people think.
Storing everything is easy. Knowing what to forget is harder. But the hardest part is when to infer versus when to accept things at face value.

[William Kearns @kearnsw]

Human memory doesn’t store transcripts. It builds models of the world and updates them over time.