Dad, fly fisherman, product builder. Founder/CEO LemonStand → exit to @Mailchimp. Recently, CEO of $TINY portco. Exploring opportunities with AI native co's.
Canada
- Tweets 6,382
- Following 2,477
- Followers 5,508
- Likes 5,152
Pinned Tweet
Many people use AI for writing.
The problem is, most of that writing sounds like AI—em dashes included!
So I built Writers Room: an open-source writing studio for Claude Code that learns how you write, what you care about, and how you edit.
(video by /brag from @ShunitHH)
🧵
12
8
2,094
This is awesome! I had to bolt on a lot of this myself with custom scripts and skills, so I'm looking forward to retiring some of it for these new agent primitives.
0.7.5 is out!
herdr is getting used more and more for multi-agent orchestration, so this release ships a native agent cli built exactly for that. more in below 👇
but first, the new agent sidebar: deeper customization with per-token colors, bold and dim, plus full filtering and sorting any way you want. from config, scripts, or plugins, all through the api!
1
166
My favourite @bot handles grocery shopping for the family. My wife is usually resistant to things like this, but she's been super happy with it.
We both iMessage it items we notice we need throughout the week. Each Sunday, it builds an order for us with those items + our staples on the cadence it learned from analyzing our previous order history.
It messages my wife with available pickup times. She picks one, and it orders for us. It's easily saved her a couple hours per week, and a whole bunch of mental load.
1
2
158
Jev seems really complementary to all the "domain specific harness" talk.
Quickly and cheaply getting probabilities or classifier decisions can feed your deterministic systems.
I'm working on a few things this would be valuable for.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
173
@bot Why did you remove the button to create a routine? Or edit an existing one? It has broken my @inkbox_ai integration, and made it impossible to do other things.
2
3
173
Hey @poteto any insight into why a button to create a new routine or edit them was totally removed?
19
I gave a fruit fly an office job.
I’m simulating the full 166,700-neuron CNS of a male fruit fly so he can perform the vital task of moving a cursor between pieces of fruit on a tiny CRT, while grotesque human-flies distract him.
25.6 million directed connections.
Science has gone too far, but I must continue.
1
183
Pretty sure I'm going to cancel my 5x. Fable usage is now next to useless and Opus 5 isn't worth it when I have SWE-2, Sol, Grok, or GLM/Kimi.
Astra is a token hog too, but it's a strong model and @thsottiaux and gang seem more interested in my business lately.
Anthropic just rug pulled every Claude Max user.
The +50% weekly limits promo is over. Replaced with +25% permanently. That is a usage CUT disguised as a perk.
30 minutes of Fable 5.1 this morning. 90% of my session limit gone. 13% of my weekly gone.
It is noticeably worse. Immediately.
This is a $200 plan. 30 minutes should not cost you 90% of anything.
Anthropic keeps finding new ways to give you less while charging the same.
2
418
thanks to @ryancarson and @dabit3 I'm trying out @DevinAI and their fusion hybrid model harness.
So far, I really like the frontier + sidekick approach.
I have it working on some internal tools for me, and it's done a great job. The CLI is pretty clean too.
1
150
Danny Halarewich retweeted
The end of flow state
You wake up, excited for the day to start.
You have a feature idea you are looking forward to implement.
You make your coffee. Sit in front of the laptop. Put on that techno set you love.
The house is empty for the next 8 hours. No one can interrupt you.
You open your IDE and begin typing.
Soon, hours pass in minutes.
Everything makes sense.
And before you know it, it’s the end of the day.
You’re looking at today’s work and are content.
You’re energised.
Because you created something you enjoy.
Fast forward.
No more.
You start a new day with an open IDE.
But, instead of pressing those keyboard buttons, you’re staring at a orange text.
“Discombulating”
“Wombanizing”
Minutes feels like hours.
Nothing makes sense.
And before you know it, it’s the end of the day.
You’re tired. Your brain hurts.
It feels like you just finished a ton of meetings.
But, at least you shipped 1,000 lines of code, right?
You should be proud.
But, the only thing you feel is hollow.
Did you create? Or did someone else?
That magical feeling from years ago, where did it go?
Rest in peace, flow state.
You were dear to me.
208
112
1,767
91,141
Depth isn’t painted on either.
Rough creek beds slow the water down. Smooth gravel lets it run. Pools show up based on the kind of water it is: farther apart on a normal run, tighter on a steep little creek, almost stacked on a cascade.
I picked a water flow that’s similar to Northwestern British Columbia. Flat runs sit at wading depth. Pools on the big river go over your head.
A handful of beaver ponds are next. Wade-deep behind a 1.2m dam, only on slow gentle creeks.
This won't be a game to cast a line + bobber anywhere and get a random bite.
You will need to read the water. To understand fish habitat. Timing. Which fly to select.
Either that, or you don't eat.
1
65
The fly fishing mechanic itself is also based on simulating the actual physics of it:
Learning to cast a fly rod well is famously hard.
Tiny errors in the stroke get amplified by the rod, then the line, then the leader. The whole thing can look fine for 80% of the motion and still collapse into a tailing loop. Sunglasses at this stage are more about not putting a hook in your eye than glare.
I’ve been building a physics sim of that system.
Rod blank under load is not the path a rigid stick would trace. The line is a long flexible body with drag. Material specs are messy and often trade secrets. The human side is multi-joint choreography. But unlike a lot of “feel” skills, a good cast has a defined, measurable target: fly leg rides over rod leg and the two never meet, loop tight, leader turns over, nothing comes near your head.
The path that prevents you from wearing an eye-patch the rest of your life is the loaded tip path. Convex is safe. Straight is the target. Concave tails every time.
Once the rod and line are simulated well enough to score that, the problem stops being “how do I cast” and becomes “what tip path produces the best scored cast, then what body motion can produce that path.”
That’s the part I’m running in the spirit of @karpathy’s autoresearch loop: generate a tip-path candidate, simulate the cast, score it, keep or discard, repeat. Thousands of permutations. Then work backwards onto a simplified 3D body holding the rod.
And all of this for a cozy video game I’m making with my son 😆
46
A creek that’s 2m wide at its source and a river that’s 26m wide at the sea come from the same rule.
The more water that drains into a point, the wider the water gets. One rule for the whole network. Nobody typed in a single width by hand.
The blue shape is the village lake: 4.4 hectares, sitting at the height where it naturally spills.
That’s home base.
1
36
Started this as a project with my son. It’s turning into an open-world wilderness game where a lost kid has to fly fish to survive.
I’m not texturing anything yet. No trees, no rocks, no grass. The land and the water have to be right first, or the fishing will be fake.
A few shots of the bones.
2
2
227
Nobody drew a single stream in this shot.
The computer finds every low spot, then walks downhill from every cell until it knows where water would actually go. What looks like a river network is just what happens when you simulate the land and let water run downhill.
2,886 stretches of water. 26 lakes and ponds. 87 km of creek and river in a 6 km square + an additional 10km that could be unlocked later. About 1.2% of the playable ground is water.
Plenty of water to explore and fish.
1
35
These mountains aren’t procedural noise with a filter on them.
Every world gets checked against real elevation data from named ranges: the Sawtooths, the Wind Rivers, the Wasatch, Paradise Valley. How steep it is. How much the land drops. How lakes sit in their bowls. How rivers actually bend.
If it can’t sit next to the real thing and pass, it fails.
This valley drops 752m and leans about 21°. Real alpine country is usually 600–1500m and a bit steeper.
Close enough. Trees come later.
1
99
Learning to cast a fly rod well is famously hard.
Tiny errors in the stroke get amplified by the rod, then the line, then the leader. The whole thing can look fine for 80% of the motion and still collapse into a tailing loop. Sunglasses at this stage are more about not putting a hook in your eye than glare.
I’ve been building a physics sim of that system.
Rod blank under load is not the path a rigid stick would trace. The line is a long flexible body with drag. Material specs are messy and often trade secrets. The human side is multi-joint choreography. But unlike a lot of “feel” skills, a good cast has a defined, measurable target: fly leg rides over rod leg and the two never meet, loop tight, leader turns over, nothing comes near your head.
The path that prevents you from wearing an eye-patch the rest of your life is the loaded tip path. Convex is safe. Straight is the target. Concave tails every time.
Once the rod and line are simulated well enough to score that, the problem stops being “how do I cast” and becomes “what tip path produces the best scored cast, then what body motion can produce that path.”
That’s the part I’m running in the spirit of @karpathy’s autoresearch loop: generate a tip-path candidate, simulate the cast, score it, keep or discard, repeat. Thousands of permutations. Then work backwards onto a simplified 3D body holding the rod.
And all of this for a cozy video game I’m making with my son 😆
1
596
I am finding that @imagine is quite good at reproducing consistent characters across videos, if you include the same reference images.
Interestingly, the style is consistent as well, despite it not having a style reference feature. If the style portion of the prompt is the same, that seems to be enough. I'm not sure if this is just a limited style palette from the model that's working in my favour.
1
136