Right?! “Here is ChatGPT’s imagination of what our food might look like” is a red flag. Nobody was using stock photos on their menus before? 🤔 Or were they
05.03.2026 21:27
👍 0
🔁 0
💬 0
📌 0
Right?! “Here is ChatGPT’s imagination of what our food might look like” is a red flag. Nobody was using stock photos on their menus before? 🤔 Or were they
When Apple Intelligence first rolled out in 2024, I had to turn off the Priority Notifications feature because of poor reasoning.
But trying it now, after all the advancements we’ve seen since then, you can see they’ve really dialed in what is urgent
Launch Now: On putting things in front of people, early and often.
allenpike.com/2026/launch-...
I claimed in 2022 that there wasn't a mathematically pure way to solve a particular graph problem, but have now added an addendum: I'm happy to report that a ML researcher at Tiktok has proven me wrong! 🧮 allenpike.com/2022/how-to-...
Thanks, Dave! I just added an addendum with info a couple of folks from Google have sent in about the bug.
A Broken Heart: Getting a 100x speedup by changing one dumb line of code
allenpike.com/2026/a-broke...
- But Google knows "allen pike" is a "person" search.
- And "allen pike" has a similar embedding to the username "apike".
- And ”apike" has a similar embedding, as a typo, to "spike".
So, Google relates "allen pike” to “Spike Jonze". And that's why AI devs need good tracing and analysis tooling 🫠
As systems include more layers of AI, it gets difficult to intuit why they behave as they do. For example, why is one of the Google results for “allen pike" the Wikipedia page for director Spike Jonze?
Zero people searching for Allen Pike are looking for info on Spike Jonze!
Despite the challenges, it's exciting to imagine that 2026 could be the year where AI becomes able to complete multi-day tasks, perform useful household robotics work, or even produce working CSS layouts where only part of the screen scrolls
TIL there's a cracked Node dev working from Camano Island (number 10 for me, so far). I feel like the different islands should have rarities and different points. For example obviously Vancouver Island is only 1 point, but if you can collect Galiano or Guemes or something, that should be worth more
One fact about the Pacific Northwest is that half of its software talent is in the cities, and the other half live on random-ass islands you've probably never heard of. The longer you live here, the more islands you'll learn that some talented developer is currently working from.
A delete confirmation dialog that forces you to type your OWN full name to confirm deletion.
It's a common pattern to make users type "DELETE", or even the name of what they're deleting, for irreversible delete actions. This forces them to think before potentially deleting something critical.
Intercom takes this to a new level, forcing you to consider: "Should I be the one deleting this?"
A Box of Many Inputs: On browser address bars, local AI, and Roger Rabbit.
allenpike.com/2025/a-box-o...
Talking to people about this, I’ve learned that people are more happy to type on their phones than I am 💫
An addendum: Why is ChatGPT for Mac So… Bad?
allenpike.com/2025/why-is-...
Why is ChatGPT for Mac so good? A look at the the different AI labs' approaches to the Mac, and the opportunities therein.
allenpike.com/2025/why-is-...
This is a fun Latent Space episode: they interview Priscilla Chan and her husband about the biology and AI work necessary to cure all diseases. www.latent.space/p/biohub
Good meetings are clear and short. Over the last couple weeks we’ve prototyped an agent that can join calls and build realtime notes – not a transcript, just clarity on key decisions.
A lot more to do, but having fun with this one 🌱 www.youtube.com/watch?v=Z2_7...
OpenAI rates gpt-5-mini 4 lightning bolts in speed
Finally tried gpt-5-mini on some tasks, hoping at "minimal" reasoning it would be faster than claude-haiku-4.5 or gpt-4.1-mini.
Nope. Two other teams said 5-mini is so slow and unreliable that they can't use it either. Apparently "4 lightning bolts" means "takes 5-10s to process 200 tokens" 💫
How to Not Get Acquired: An approach for minimizing distraction when trying to do something rather hard. allenpike.com/2025/how-not...
The most important part of this post-mortem is crammed into a dense paragraph as long as a giraffe's vocal cord
While the AWS team has resolved the severe us-east-1 outages, reading this post-mortem makes one thing clear: they are still suffering from a critical shortage of paragraph breaks
I would have thought from M2 to M5 you could have gotten the same performance with less chassis but apparently that was not the goal
“When we released Vision Pro, it provided glimpse of what spatial computing can be. Consistently, though, we’ve heard your concerns about weight and comfort. That’s why we’ve made the M5 Vision Pro 25% heavier – ensuring it is pressing on your nose, cheeks, and face with unprecedented heft”
UX Entropy: How products like Zoom can bloat from hero to hulk.
allenpike.com/2025/ux-entr...
Building Something Big: On the pros and cons of the indie path. allenpike.com/2025/buildin...
This makes me wonder – why doesn't the operating system do this? Like if Slack is on my left screen and a notification comes in, it would be nice to get the sound coming "from" Slack, if I'm wearing stereo headphones?
Cursor's UI is nicer than using Claude Code in a terminal, but one helpful thing about Claude is that you can script hooks. For example, with an agent going on each screen, you can have Claude chime where it needs review or approval – panned left for the left screen, or right for the right. 🔊
Every answer was "Let's be real: you want to know how tall giraffes are, but why are you even asking? We need to go to verified sources – then go BEYOND giraffes and assess all tall mammals. The big picture, man."
ChatGPT trait hint: "You don't need to compliment me. Adopt the tone of critical thinking, and be willing to revisit and question assumptions. Push me or challenge me to do better. When something might not be true, equivocate - for example, use phrases like “might be” or “is probably”, or mention relevant tradeoffs. In voice mode, speak quickly."
If you're getting weird behaviour from ChatGPT 5, check "What traits should ChatGPT have?" under Customize ChatGPT. I had this prompt there to help 4o be more critical, but they trained GPT-5 to be more critical by default, so the prompt pushed it into overcaffeinated paranoia.
Getting Tied Up: On little things that get in the way.
allenpike.com/2025/getting...