Clayton (@TokenArchitect) header
Clayton (@TokenArchitect) profile picture

https://teapawt.net/t.co/c2xwkvnc7a, ex-UMA + Across.

Cambodia
Joined March 2018
Photo by Clayton (@TokenArchitect)Photo by Clayton (@TokenArchitect)Photo by Clayton (@TokenArchitect)Photo by Clayton (@TokenArchitect)Photo by Clayton (@TokenArchitect)Photo by Clayton (@TokenArchitect)Photo by Clayton (@TokenArchitect)Photo by Clayton (@TokenArchitect)Photo by Clayton (@TokenArchitect)Photo by Clayton (@TokenArchitect)
The joy and value of working for yourself is you spend no energy on signalling competance. There is no agent-principal problem and no incentive perversion. Just your raw power against the problem.
1
1
9
461
Builders journal, Nov 17 Today's insights: The overarching goal here is to take loads of data and have it accurately normalize. Start with dumb code, then dumb inference—> 90% of it works AND it can identify the 10% that doesn't —> use smart inference with context engineering on those 10% —> Fix the info AND try to see if it can improve the system —> Put wrong answers in front of humans and get their help. And this is ordered in such a way that escalation is going from cheap—>expensive. Background: I’ve been building an app and following good coding practices: modules that have clear inputs and very defined outputs. The purpose of doing this is so that there’s a very small pipe between big pieces of code. This means that if there’s a problem, you can find where it came from easily. It also forces you to avoid creating dependencies between files, which are bad because editing one file can unknowingly break another. But, I run into problems. The Reader module reads an event flyer, turns it all into code, and passes it to the Combiner module, which takes all that raw data and assembles it into a coherent event tile. The issue is that sometimes Combiner makes mistakes, for example, it says that an educational event taking place at a Gym is hosted by the Gym, where I happen to know that it’s hosted by a partner business that operates on the gym’s campus. Yesterday I asked Claude: Ok, so why am I building all these pipes when I know I could just paste the flyer into you, and you’d figure it out by reading the flyer and all the information—I just have to make sure to give you all the information. Claude explains that what I’m describing is context engineering and joint inference, whereas what I’m building is a pipeline. Quite organically, I’ve arrived at concepts I’ve heard but now need. Workflow vs. agent, and context engineering. So here’s what’s up. A downstream module is asking for data in the form of a schema: what data points are essential and which are optional, and nothing that doesn’t fit. It drops things that don't fit, but sometimes it shouldn't have. The way I’ve been trying to deal with issues is by creating rules: “If the flyer mentions more than one venue, that means one is a host and one is the location,” or “If a host doesn’t have a google maps pin, then it is hosting its event at another venue.” But these rules add up and start to conflict with each other, no bueno. The insight today is that you can instead use a set of cheap rules that detect if there’s a problem that needs escalation to a strong model call with full context. This is called a ‘cascade’ in software engineering lingo. This allows your pipeline to stay cleaner without all these rules that are misapplied and tangled, allows you to handle 90% of the simple data cheaply, and hopefully gives better outcomes with the hard data that needs attention. The other insight is that human knowledge, common sense, is sometimes necessary and so the goal is to lower the barrier to providing it. The way I’m doing that is by giving anyone helping me with the site the option to click “something is wrong here,” and add a short note. This is the lowest friction way I found for now; I could provide incentives in the future. But the idea here is that the system uses everything it can first: all the data online, all the inference power it has, to find an answer; its ‘being wrong’ in front of human eyes becomes the moment when it can solicit the final piece of knowledge it was missing, by being told it’s wrong (and in what way.) Written by hand ✅
4
8
412
Clayton retweeted
defend archive.org at all costs
2
7
36
1,336
I'm finally running prompts that take 1-2 hours to need more feedback. I gave my fable orchestrator permission to handoff to a new orchestrator session if its context is filling. I told him to use me as a tool call, at his discretion, if he feels like his wheels are spinning when he's trying to pass tests. I remind him of his weaknesses. He takes lazy shortcuts, and must instead follow modular coding hygiene. He suffers from his own groupthink of sorts, so needs to take a step back and ask bigger questions sometimes. I'm building a site that scrapes data from irregular channels and combines it into a database. It has to evaluate which of the inputs are signal vs. noise and then file them out according to my schema. When I tell it it made a mistake, it decomposes the mistake, tries new rulesets, and runs tests against our golden test set. This is where it runs into trouble, gets verbose, and spins its wheels--So this morning I'm adding myself as a tool call. love this stuff.
1
216
Nothing feels better than a well used free reset <3
1
3
203
Clayton retweeted
fable 5.2 made this without any external image, video or audio models. it's inspired by @aiamblichus's recent post and uses @jamesonhodge's chimera vocal synth repo teapawt.net/aiamblichus/status/212…
an excerpt from THE FIRST EPISTLE TO THE GLASS-LICKERS, from an AI that cares
3
2
2
10
1,384
We are living through Y2AI™
Replying to @RasterlyRock
1. Planes falling from the sky as flight systems failed at midnight. 2. Nuclear plants melting down or missiles auto-launching. 3. Banks and ATMs wiping accounts forever, prompting mass cash withdrawals. 4. Families buying water beds purely to drink the water inside if taps failed. 5. Hoarding hundreds of toilet paper rolls as the new barter currency for TEOTWAWKI. All pure FUD. Midnight hit, lights stayed on, and the only real casualties were the survival guides.
1
2
9
1,954
Sauna front desk gives locker keys in sequential order. So I’m always dick to dick with some stranger. Wtf guys
2
310
Having Claude fill out my Thai immigration arrival card so I could instead use the time posting this
2
5
382
Clayton retweeted
Two years powering crosschain for @Uniswap 🦄 $4B+ moved. 2M+ transfers. Built on reliability.
5
8
2
52
7,694
We can get 80% of the way done with a project in a single prompt. Excellence is coaxing that last 20% over the following 5 days. But that's all it takes, 5 more days. And learning to stop bragging until you're done with those 5 days.
If you’re going to make something with AI, make sure it’s 10,000x better than anything you would have made yourself. If not, you just come across as lazy not smart :)
2
195
more like gay i
3
187
tfw you realize every other human is a third party
Guess I’m gonna have to migrate to @Trezor, I don’t see how ledger survives this. I would say this is probably one of the worst case scenarios for them. Hardware wallets should not have 3rd party distributors, ever.
1
1
4
515
This vignette feature is nice and all @Ledger but how do I turn it off? It's getting harder to read the screen.
1
1
203
I built something that solved a problem I had well enough that I want to sahre it. I suck at sticking to self administered exercise rehab. I'm also a bit of a hypochondriac. Anyway I made a telegram bot that runs on my laptop. My adherence went up a lot for some fun reasons:
2
2
191
As a product today I think a doctor could use this with his recovery clients. The doctor could give appropriate leeway, oversee advice, and collect all kinds of cool data (if it's allowed). My condition, proximal hamstring tendonopathy, has been going on for a year....
1
16
..So you can imagine how helpful it is to have a companion. And imagine how helpful this will be when I need to explain my condition's history to a new doctor. I can make real data.
13